A lightweight and generalizable deep learning framework for early detection of rice leaf diseases in complex field environments Authors: Chong Zhang, Xiaoxi Hao, Jianan Liang, Lili Li, Wenwei Li Published: 2025 ABSTRACT Rice leaf diseases pose a significant and escalating threat to global food security. Timely and accurate detection, particularly in the critical early stages characterized by subtle lesions, is paramount for effective disease management. However, existing solutions often struggle with the complexities of real-world field environments (e.g., variable lighting, occlusions, complex backgrounds), computational constraints on edge devices, and limited generalizability across diverse disease types and plant species. To address these challenges, this study proposes a novel lightweight deep learning framework specifically designed for robust rice leaf disease detection. Our key innovations include: (1) A Multi-branch Large-kernel Fusion Depthwise (MLFD) module enhancing multi-scale contextual feature extraction critical for identifying subtle early lesions; (2) A Multi-scale Dilated Transformer Attention (MDTA) module integrating spatial and channel attention mechanisms to improve feature representation under complex conditions; (3) A Lightweight Detection Head (Lo-Head) optimized with grouped and depthwise convolutions, drastically reducing model complexity without sacrificing accuracy. Crucially, extensive experiments demonstrate the framework’s superior performance. On a dedicated rice leaf disease dataset, it achieves a mean Average Precision mAP@0.5:0.95 of 62.62%, outperforming state-of-the-art lightweight detectors including YOLOv5n (56.73%), YOLOv8n (57.41%), YOLOv10n (56.14%), and the baseline YOLOv11n (60.85%), while maintaining low computational demands (6.3 GFLOPs, 2.66M parameters). Significantly, rigorous generalization experiments validate the model’s exceptional transferability. Evaluated on independent datasets encompassing potato and tomato leaf diseases, the proposed framework consistently surpasses comparable models in mAP@0.5:0.95, demonstrating its robust capability to detect diseases across different plant species. This combination of high accuracy, computational efficiency, and remarkable cross-crop generalizability positions our framework as a highly promising tool for practical deployment on resource-limited edge devices (e.g., drones, field sensors) in smart agriculture systems, enabling proactive disease surveillance and precision control strategies across diverse crops. Introduction Rice is an essential staple food globally, supporting vast populations and remaining central to agricultural systems across many Asian nations. Its high yield per unit area allows for efficient land resource utilization. Consequently, safeguarding rice health is synonymous with preserving food security, enhancing land-use economics, and contributing to ecological stability1,2. For many decades, China has been a primary contributor to global rice output and domestic consumption. Despite this, rice production is increasingly jeopardized by shifting climates, changes in cultivation practices, and the proliferation of pests and diseases3. Specifically, leaf diseases–including rice blast, brown spot, bacterial streak, and blight–are highly destructive and represent a critical menace to food supplies4. Rice blast, a fungal disease, is considered the most intricate and destructive rice pathology due to its multi-site infestation and ability to persist across the entire growth cycle5. Its symptoms include diamond, oval, or spindle-shaped leaf spots featuring gray/white centers and brown borders. Severe outbreaks can cause total leaf desiccation, compromising photosynthesis and growth, which precipitates significant yield reduction. Similarly, brown spot manifests as small, oval lesions, initially light brown but darkening centrally, often surrounded by a yellow halo, resulting in localized leaf tissue necrosis6. Bacterial leaf streak, a bacterial infection, thrives in warm, humid conditions with ample rainfall. The pathogen invades leaves to form small, dark green, translucent spots that elongate into yellowish-brown streaks confined by the leaf veins7. Finally, bacterial leaf blight, caused by the bacterium Xanthomonas oryzae pv. oryzae, typically enters via stomata or marginal wounds. This disease is characterized by initial small, water-soaked, yellow-green lesions along leaf margins that rapidly lengthen along the veins, forming long, wavy, yellow-to-white stripes. Coalescence of these lesions causes the leaf to dry out completely. Advanced infection stages result in premature senescence, diminished photosynthetic capacity, and severe crop loss8. The reliance of traditional disease detection techniques on manual expertise renders them time-consuming, inherently subjective, and inadequate for the rapid scaling requirements of modern agriculture. Consequently, developing highly efficient and accurate detection technologies is paramount for enhancing disease control capabilities and securing food supplies9. Swift diagnosis and accurate localization of rice diseases are fundamental requirements for effective preventative management. In this context, image-based object detection technology has gained substantial traction as a research focus, offering solutions that are efficient, objective, and scalable10. By integrating high-resolution imaging (often from cameras or drones) with advanced deep learning algorithms, accurate disease identification and spatial localization can be achieved11. Agricultural disease detection research generally employs two categories of deep learning models. The first consists of two-stage detection frameworks, such as Faster R-CNN12 and Mask-RCNN13. These models prioritize detection precision by utilizing a region proposal stage before final classification. The second category comprises one-stage, regression-based networks like YOLO14 and SSD15, which are highly valued for their exceptional real-time inference speed. Over recent iterations, the YOLO network specifically has seen improvements that boost both detection accuracy and processing speed, leading to its widespread application in industrial defect analysis and large-scale agricultural monitoring of crops and fruit16. Related work Current research has demonstrated considerable advancements in automated plant leaf disease detection. For instance, Ramadan et al.17 applied Generative Adversarial Networks (GANs) to augment rice leaf datasets, successfully increasing data volume and diversity. This strategy mitigated overfitting risks and consequently boosted the overall detection performance. Tao et al.18 introduced CEFW-YOLO, a lightweight network built upon YOLOv11n, which integrated channel compression convolution and a cross-channel attention mechanism. This architectural change significantly strengthened the network’s capacity for global feature capture, leading to improvements in both accuracy and operational robustness. Wang et al.19 focused on enhancing the YOLOv8 framework, achieving a mean Average Precision (mAP) of 94.8%, which enabled the detection and subsequent grading of fresh tea tree leaves. Xu et al.20 designed the novel SPLDPvB network, which proved capable of precisely identifying leaf spot disease and common pests in sweet potatoes. In a different crop context, Zhou et al.21 developed a high-precision methodology for detecting gray spot disease in apples. Furthermore, Shi et al.22 improved field-level cabbage seedling detection through the incorporation of the CBAM23 and BiFPN modules24. Abulizi et al.25 enhanced the YOLOv9 structure by integrating DySample26 and MPDIoU27, aiming to improve the feature extraction for small lesions, refine boundary accuracy, and reduce background interference. Focusing on feature granularity, Zhou et al.28 introduced cross-modal retrieval for rice leaf lesion detection by employing ViT29 and BERT30 for fine-grained feature extraction. Their model’s learning efficiency was further optimized via a bidirectional hybrid self-attention mechanism and a warm-up bat algorithm for dynamic learning rate scheduling31. Finally, Deari et al.32 demonstrated robust rice leaf disease classification by fusing classification and detection networks, an approach that effectively minimized the occurrence of false positives. In parallel with advancing model accuracy, the topic of Explainable Artificial Intelligence (XAI) in deep learning for agriculture has gained increasing prominence. Understanding why a model makes a specific prediction is crucial for building trust and reliability, particularly in real-world applications where decision errors can be costly. Research efforts have explored various techniques to interpret complex deep neural networks. For instance, studies have employed attention mechanisms and feature analysis to enhance model interpretability in crop classification tasks33,34. Other works have integrated hybrid feature aggregation networks with specialized encoding attention mechanisms to improve both classification performance and the interpretability of disease identification in crops like cassava35. Furthermore, the application of ensemble learning alongside XAI techniques has been investigated to enhance the reliability and transparency of plant leaf disease detection systems36. These approaches highlight the growing consensus that interpretability is as essential as raw performance in modern agricultural AI systems. Limitations of existing studies Despite the significant progress in plant leaf disease detection, several critical challenges remain for practical deployment in real-world agricultural settings. Specifically, existing methods face the following bottlenecks, which motivate the current research:Limited Robustness under Complex Field Conditions: Many models achieve high accuracy on clean, laboratory-like datasets but exhibit performance degradation in real field environments characterized by variations in illumination, complex backgrounds (soil, shadows, other plants), and different growth stages (e.g37.,). Addressing inter- and intra-class variation under complex field conditions remains a major hurdle.High Computational Costs: State-of-the-art high-accuracy models often have large parameter counts and require substantial computational resources (e.g38.,). This high demand hinders their real-time deployment and inference capabilities on resource-constrained edge devices (such as smart sensors or low-power drones) commonly used in precision agriculture.Poor Generalizability and Transferability: Specialized detection models, while achieving good accuracy for a specific disease or crop (e.g39.,), often lack generalizability. They require complete retraining for new crop species or different disease types, preventing the development of a unified and cost-effective solution for multi-crop monitoring.Difficulty in Detecting Subtle Early Lesions: Early-stage disease spots are typically small and subtle, making multi-scale feature extraction challenging. Many existing architectures struggle to effectively capture the fine-grained, low-level features necessary for early diagnosis. Limited Robustness under Complex Field Conditions: Many models achieve high accuracy on clean, laboratory-like datasets but exhibit performance degradation in real field environments characterized by variations in illumination, complex backgrounds (soil, shadows, other plants), and different growth stages (e.g37.,). Addressing inter- and intra-class variation under complex field conditions remains a major hurdle. Limited Robustness under Complex Field Conditions: Many models achieve high accuracy on clean, laboratory-like datasets but exhibit performance degradation in real field environments characterized by variations in illumination, complex backgrounds (soil, shadows, other plants), and different growth stages (e.g37.,). Addressing inter- and intra-class variation under complex field conditions remains a major hurdle. High Computational Costs: State-of-the-art high-accuracy models often have large parameter counts and require substantial computational resources (e.g38.,). This high demand hinders their real-time deployment and inference capabilities on resource-constrained edge devices (such as smart sensors or low-power drones) commonly used in precision agriculture. High Computational Costs: State-of-the-art high-accuracy models often have large parameter counts and require substantial computational resources (e.g38.,). This high demand hinders their real-time deployment and inference capabilities on resource-constrained edge devices (such as smart sensors or low-power drones) commonly used in precision agriculture. Poor Generalizability and Transferability: Specialized detection models, while achieving good accuracy for a specific disease or crop (e.g39.,), often lack generalizability. They require complete retraining for new crop species or different disease types, preventing the development of a unified and cost-effective solution for multi-crop monitoring. Poor Generalizability and Transferability: Specialized detection models, while achieving good accuracy for a specific disease or crop (e.g39.,), often lack generalizability. They require complete retraining for new crop species or different disease types, preventing the development of a unified and cost-effective solution for multi-crop monitoring. Difficulty in Detecting Subtle Early Lesions: Early-stage disease spots are typically small and subtle, making multi-scale feature extraction challenging. Many existing architectures struggle to effectively capture the fine-grained, low-level features necessary for early diagnosis. Difficulty in Detecting Subtle Early Lesions: Early-stage disease spots are typically small and subtle, making multi-scale feature extraction challenging. Many existing architectures struggle to effectively capture the fine-grained, low-level features necessary for early diagnosis. Research contributions To address the aforementioned limitations, we propose a novel lightweight and generalizable deep learning framework specifically engineered for early rice disease detection in complex field environments. Our core contributions are explicitly outlined below:Multi-branch Large-kernel Fusion Depthwise (MLFD) module: We propose a novel MLFD module that replaces standard convolutions. It utilizes parallel large-kernel depthwise convolutions to significantly enhance multi-scale feature extraction for subtle early lesions, improving the model’s ability to detect small and faint spots while maintaining computational efficiency.Multi-scale Dilated Transformer Attention (MDTA) module: We introduce the MDTA module, which integrates spatial-channel attention mechanisms with dilated convolutions. This module effectively expands the receptive field without increasing parameters, improving feature representation and enhancing the model’s robustness against environmental noise (e.g., shadows, light variation) and background clutter.Lightweight Detection Head (Lo-Head): We design an optimized Lightweight Detection Head (Lo-Head) that utilizes grouped convolutions. This optimization drastically reduces the number of trainable parameters, enabling high-speed, real-time inference on resource-limited edge devices for practical agricultural deployment.Demonstrated Cross-Species Generalizability and Superior Performance: Beyond validating our model’s effectiveness on rice, we conducted a systematic evaluation of its cross-species generalizability. When training our proposed method on both potato and tomato leaf disease datasets, our approach achieved a significant improvement in detection accuracy compared to the baseline model on both new crops. This result confirms the framework’s robust cross-crop versatility and superior performance, offering an efficient solution for unified multi-crop disease monitoring. Multi-branch Large-kernel Fusion Depthwise (MLFD) module: We propose a novel MLFD module that replaces standard convolutions. It utilizes parallel large-kernel depthwise convolutions to significantly enhance multi-scale feature extraction for subtle early lesions, improving the model’s ability to detect small and faint spots while maintaining computational efficiency. Multi-branch Large-kernel Fusion Depthwise (MLFD) module: We propose a novel MLFD module that replaces standard convolutions. It utilizes parallel large-kernel depthwise convolutions to significantly enhance multi-scale feature extraction for subtle early lesions, improving the model’s ability to detect small and faint spots while maintaining computational efficiency. Multi-scale Dilated Transformer Attention (MDTA) module: We introduce the MDTA module, which integrates spatial-channel attention mechanisms with dilated convolutions. This module effectively expands the receptive field without increasing parameters, improving feature representation and enhancing the model’s robustness against environmental noise (e.g., shadows, light variation) and background clutter. Multi-scale Dilated Transformer Attention (MDTA) module: We introduce the MDTA module, which integrates spatial-channel attention mechanisms with dilated convolutions. This module effectively expands the receptive field without increasing parameters, improving feature representation and enhancing the model’s robustness against environmental noise (e.g., shadows, light variation) and background clutter. Lightweight Detection Head (Lo-Head): We design an optimized Lightweight Detection Head (Lo-Head) that utilizes grouped convolutions. This optimization drastically reduces the number of trainable parameters, enabling high-speed, real-time inference on resource-limited edge devices for practical agricultural deployment. Lightweight Detection Head (Lo-Head): We design an optimized Lightweight Detection Head (Lo-Head) that utilizes grouped convolutions. This optimization drastically reduces the number of trainable parameters, enabling high-speed, real-time inference on resource-limited edge devices for practical agricultural deployment. Demonstrated Cross-Species Generalizability and Superior Performance: Beyond validating our model’s effectiveness on rice, we conducted a systematic evaluation of its cross-species generalizability. When training our proposed method on both potato and tomato leaf disease datasets, our approach achieved a significant improvement in detection accuracy compared to the baseline model on both new crops. This result confirms the framework’s robust cross-crop versatility and superior performance, offering an efficient solution for unified multi-crop disease monitoring. Demonstrated Cross-Species Generalizability and Superior Performance: Beyond validating our model’s effectiveness on rice, we conducted a systematic evaluation of its cross-species generalizability. When training our proposed method on both potato and tomato leaf disease datasets, our approach achieved a significant improvement in detection accuracy compared to the baseline model on both new crops. This result confirms the framework’s robust cross-crop versatility and superior performance, offering an efficient solution for unified multi-crop disease monitoring. Materials and methods DataSets Table 1Detailed construction of the Rice Leaf Disease Detection Dataset (4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} Augmentation). Dataset type Disease class Original imagesTraditional Aug.CycleGAN Aug.Augmented totalTrain Set(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_o$$\end{document}=4426,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}=17704)Brown Spot (BS)1320198019805280Leaf Blast (LB)1320198019805280Bacterial Leaf Blight (BLB)893134013403573Bacterial Leaf Streak (BLS)893134013383571Subtotal44266640663817704Validation Set(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V_o$$\end{document}=1265)Brown Spot (BS)378--378Leaf Blast (LB)378--378Bacterial Leaf Blight (BLB)255--255Bacterial Leaf Streak (BLS)254--254Subtotal1265--1265Test Set(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_o$$\end{document}=632)Brown Spot (BS)189--189Leaf Blast (LB)189--189Bacterial Leaf Blight (BLB)128--128Bacterial Leaf Streak (BLS)126--126Subtotal632--632 Final Total 6323 6640 6638 19601 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}represents the number of samples in the augmented dataset. Data Sources: The dataset utilized in this study was curated and screened from the following publicly available online repositories:.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://www.kaggle.com/datasets/douaairaq/rice-disease-dataset-labelled-to-train-yolov5.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://universe.roboflow.com/tip-kwadd/group5-rice-leaf-disease-datasets/dataset/3.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://universe.roboflow.com/tru-projects-cqcql/plantdet-rice-leaf-disease-setection/dataset/1. Table 1Detailed construction of the Rice Leaf Disease Detection Dataset (4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} Augmentation). Dataset type Disease class Original imagesTraditional Aug.CycleGAN Aug.Augmented totalTrain Set(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_o$$\end{document}=4426,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}=17704)Brown Spot (BS)1320198019805280Leaf Blast (LB)1320198019805280Bacterial Leaf Blight (BLB)893134013403573Bacterial Leaf Streak (BLS)893134013383571Subtotal44266640663817704Validation Set(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V_o$$\end{document}=1265)Brown Spot (BS)378--378Leaf Blast (LB)378--378Bacterial Leaf Blight (BLB)255--255Bacterial Leaf Streak (BLS)254--254Subtotal1265--1265Test Set(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_o$$\end{document}=632)Brown Spot (BS)189--189Leaf Blast (LB)189--189Bacterial Leaf Blight (BLB)128--128Bacterial Leaf Streak (BLS)126--126Subtotal632--632 Final Total 6323 6640 6638 19601 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}represents the number of samples in the augmented dataset. Data Sources: The dataset utilized in this study was curated and screened from the following publicly available online repositories:.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://www.kaggle.com/datasets/douaairaq/rice-disease-dataset-labelled-to-train-yolov5.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://universe.roboflow.com/tip-kwadd/group5-rice-leaf-disease-datasets/dataset/3.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://universe.roboflow.com/tru-projects-cqcql/plantdet-rice-leaf-disease-setection/dataset/1. Detailed construction of the Rice Leaf Disease Detection Dataset (4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} Augmentation). Train Set (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_o$$\end{document}=4426,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}=17704) Validation Set (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V_o$$\end{document}=1265) Test Set (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_o$$\end{document}=632) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}represents the number of samples in the augmented dataset. Data Sources: The dataset utilized in this study was curated and screened from the following publicly available online repositories:. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://www.kaggle.com/datasets/douaairaq/rice-disease-dataset-labelled-to-train-yolov5. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://universe.roboflow.com/tip-kwadd/group5-rice-leaf-disease-datasets/dataset/3. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bullet$$\end{document}https://universe.roboflow.com/tru-projects-cqcql/plantdet-rice-leaf-disease-setection/dataset/1. Fig.1Representative Early-Stage Rice Leaf Disease Examples. This figure presents typical visual characteristics of four common early-stage rice leaf diseases. Fig. 2Examples of the Rice Leaf Disease Dataset with Augmentation Strategies. The figure illustrates a selection of images from the constructed rice leaf disease dataset. The top row, “Original Images,” displays raw samples. The middle row, “Traditional Enhanced Images,” shows examples processed with traditional augmentation techniques (e.g., rotation, flipping, brightness adjustments). The bottom row, “CycleGAN Enhanced Images,” presents synthetic images generated by the CycleGAN model, showcasing enhanced diversity and realism for training. Representative Early-Stage Rice Leaf Disease Examples. This figure presents typical visual characteristics of four common early-stage rice leaf diseases. Examples of the Rice Leaf Disease Dataset with Augmentation Strategies. The figure illustrates a selection of images from the constructed rice leaf disease dataset. The top row, “Original Images,” displays raw samples. The middle row, “Traditional Enhanced Images,” shows examples processed with traditional augmentation techniques (e.g., rotation, flipping, brightness adjustments). The bottom row, “CycleGAN Enhanced Images,” presents synthetic images generated by the CycleGAN model, showcasing enhanced diversity and realism for training. The dataset utilized in this study for rice leaf disease detection comprises \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6323}$$\end{document} original images covering four primary disease classes. These classes represent typical early stage rice leaf diseases, as illustrated in Fig. 1. The complete dataset was constructed from publicly available online resources, and detailed citations are provided in Table 1. This original dataset size, however, presents a significant challenge for training complex deep neural networks, notably increasing the risk of overfitting and constraining the model’s capacity to generalize across the substantial visual variability observed in real-world agricultural environments. Factors such as lighting conditions, occlusions, and diverse disease manifestations contribute to this complexity. To effectively mitigate these limitations and enhance model robustness, this study employed a comprehensive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {4}\times$$\end{document} data augmentation strategy applied solely to the training set. This strategy involved a combination of traditional image processing techniques and advanced synthetic image generation. Specifically, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6640}$$\end{document} images were generated using traditional augmentation (including geometric and photometric transformations), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6638}$$\end{document} images were synthesized using a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {CycleGAN}$$\end{document}-based approach to enrich both the realism and diversity of the training data. This augmentation process expanded the original \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {4426}$$\end{document} training samples (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_o$$\end{document}) into a total training set of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {17704}$$\end{document} images (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_a$$\end{document}).All images across the final dataset were normalized to a uniform resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {640} \times \text {640}$$\end{document} pixels to ensure compatibility with the network input structure. The complete dataset, totaling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {19601}$$\end{document} samples, was subsequently partitioned into training, validation, and test sets using a strict \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7:2:1}$$\end{document} ratio, respectively, ensuring mutual exclusivity. A representative example of the augmented images is illustrated in Fig. 2. General technical route Fig. 3Overall Framework for the Development and Iterative Optimization of the Proposed Rice Leaf Disease Detection Network. Overall Framework for the Development and Iterative Optimization of the Proposed Rice Leaf Disease Detection Network. This study systematically details the development of a rice leaf disease detection network built upon the enhanced YOLOv11 architecture, adhering to established protocols for deep learning model engineering. Our research commenced with the Data Preparation phase, aiming to provide stable and high-quality input for the subsequent model training process through strict and comprehensive processing. In this stage, we performed key operations such as Data Cleaning, Image Preprocessing, Data Partitioning, and Data Augmentation. Subsequently, the process enters the Model Optimization/Performance Evaluation phase, where the improved network architecture is defined through Model Configuration, and efficient training algorithms are utilized to accurately learn disease characteristics. The key in this phase lies in Model Training and Hyperparameter Tuning, which, combined with Testing and Validation and Performance Evaluation, form a tight feedback loop to ensure optimal model performance. Finally, the rigorously evaluated model is prepared for Final Model Deployment. The stages of the entire research process are closely connected and iteratively optimized. The trained model is ultimately applied to the test data for inference and validation to rigorously examine its effectiveness in real-world scenarios. The detailed flow is shown in Fig. 3. The proposed framework Fig. 4Detailed Architecture of the Proposed Lightweight Detection Network. The network architecture is composed of three main components: the Backbone, the Neck, and the Head. The Backbone is responsible for feature extraction, integrating the Spatial Pyramid Pooling Fusion (SPPF) module and the Multi-scale Dilated Transformer Attention (MDTA) module, along with the Multi-branch Large-kernel Fusion Depthwise (MLFD) module. The Neck utilizes Upsample operations and MLFD blocks for effective feature fusion across different scales. The Head employs the Lightweight Detection Head (Lo-Head) to predict bounding boxes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Bbox}$$\end{document}) and classification probabilities (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {cls}$$\end{document}), outputting logits for the Distribution Focal Loss (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {DFL}$$\end{document}) and Binary Cross-Entropy Loss (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {B-loss}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {C-loss}$$\end{document}). The detailed internal structures of the MLFD, MDTA, and Lo-Head modules are presented in the dashed boxes above. Detailed Architecture of the Proposed Lightweight Detection Network. The network architecture is composed of three main components: the Backbone, the Neck, and the Head. The Backbone is responsible for feature extraction, integrating the Spatial Pyramid Pooling Fusion (SPPF) module and the Multi-scale Dilated Transformer Attention (MDTA) module, along with the Multi-branch Large-kernel Fusion Depthwise (MLFD) module. The Neck utilizes Upsample operations and MLFD blocks for effective feature fusion across different scales. The Head employs the Lightweight Detection Head (Lo-Head) to predict bounding boxes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Bbox}$$\end{document}) and classification probabilities (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {cls}$$\end{document}), outputting logits for the Distribution Focal Loss (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {DFL}$$\end{document}) and Binary Cross-Entropy Loss (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {B-loss}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {C-loss}$$\end{document}). The detailed internal structures of the MLFD, MDTA, and Lo-Head modules are presented in the dashed boxes above. The complexity of early-stage rice leaf disease detection, characterized by subtle, fine-grained lesions in dynamic field environments (variable light, shadows), necessitates a highly efficient yet robust architecture. Our proposed framework introduces three specialized modules integrated into the YOLOv11n backbone: MLFD, MDTA, and Lo-Head. First, we replace the original C3k2 component with the novel Multi-branch Large-kernel Fusion Depthwise (MLFD) module. This module is designed as the primary, computationally-light backbone feature extractor. Its architecture employs a parallel multi-branch topology and integrates large-kernel depthwise separable convolutions, which is motivated by the need to efficiently capture the subtle texture and extended context of early lesions while mitigating the trade-off between receptive field size and computational cost. To refine contextual feature representation and address the known limitations of standard channel-wise attention within multi-scale feature hierarchies, we introduce the Multi-scale Dilated Transformer Attention (MDTA) module. MDTA is strategically incorporated into the network’s neck to bolster feature resilience against complicated backgrounds and varying target scales. It achieves this by combining a Transformer-inspired attention mechanism with spatial dilation. This structural choice grants the model an exponentially expanded receptive field without introducing additional parameters, enabling the capture of vital long-range dependencies and non-local context. This capability is key to effectively discriminating small lesions from surrounding healthy tissue or environmental noise. Within the MDTA structure, CA signifies Channel Attention, while SA denotes Spatial Attention, highlighting its dual-path architecture for comprehensive feature refinement. Subsequently, the conventional detection head is substituted with a lightweight architectural variant termed Lo-Head. This modification is motivated by the critical need for real-time inference capability on edge devices with restricted resources. The Lo-Head drastically minimizes both the parameter count and floating-point operations (FLOPs), primarily by utilizing group convolutions and an optimized decoupled structure. This ensures the resultant final architecture is fully suitable for deployment on agricultural platforms demanding high Frames Per Second (FPS) performance. The complete structural design of the proposed framework, illustrating the seamless integration of the MLFD, MDTA, and Lo-Head modules into the overall detection pipeline, is presented in Fig. 4. MLFD: Multi-branch large-kernel fusion depthwise module Fig. 5Detailed Architecture of the Multi-branch Large-kernel Fusion Depthwise (MLFD) Module. This figure elucidates the comprehensive design of the MLFD module. Subfigure (a) illustrates the multi-branch structure, showing how features from different convolutional paths are integrated via element-wise sum, contributing to enhanced multi-scale feature extraction. Subfigure (b) details the cascade feature fusion mechanism, where features from different stages of the module are concatenated and further processed, enriching contextual information. Subfigure (c) presents the internal architecture of the DepthBottleneckUniv2, a key component within MLFD, which employs Dilated ReparamBlock, BatchNorm, and UniRepLKNetBlock for efficient and robust feature processing. Detailed Architecture of the Multi-branch Large-kernel Fusion Depthwise (MLFD) Module. This figure elucidates the comprehensive design of the MLFD module. Subfigure (a) illustrates the multi-branch structure, showing how features from different convolutional paths are integrated via element-wise sum, contributing to enhanced multi-scale feature extraction. Subfigure (b) details the cascade feature fusion mechanism, where features from different stages of the module are concatenated and further processed, enriching contextual information. Subfigure (c) presents the internal architecture of the DepthBottleneckUniv2, a key component within MLFD, which employs Dilated ReparamBlock, BatchNorm, and UniRepLKNetBlock for efficient and robust feature processing. In the task of disease detection on rice leaves, a lightweight and efficient detection network is essential to strike a balance between speed and accuracy for real-time applications. The C3k2 module in the original network, while capable of combining different convolutional kernels for multi-scale features, inherently increases computational complexity, affecting real-time performance. Furthermore, its capacity to manage variations from shading, complex backgrounds, and capture fine details of very small targets is often insufficient for the diverse and complex data found in natural field environments. This makes the original C3k2 module weak in early disease detection of rice leaves. The MLFD module is engineered to amplify feature diversity and enrich semantic representation while maintaining minimal computational overhead. As depicted in Fig. 5, its architectural foundation is a parallel, multi-branch topology (refer to subfigure (a)), which processes the incoming feature map simultaneously across several independent pathways. This inherent parallelism is vital for effectively capturing lesion characteristics across diverse scales and orientations, thereby enhancing the network’s expressive flexibility. Each pathway incorporates multiple DepthBottleneckUniv2 units. These units leverage large-kernel operations (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times 5$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7\times 7$$\end{document}) to efficiently expand the receptive field, a necessary step for integrating crucial contextual information surrounding subtle, incipient lesions. To foster inter-branch synergy, the MLFD module implements a cascaded fusion strategy, utilizing ELAN-style feature concatenation to merge the diverse representations (as illustrated in subfigure (b)). This mechanism promotes optimal information flow and ensures the network learns more discriminative and potent features. The internal design of the DepthBottleneckUniv2 unit–detailed in subfigure (c)–is constructed from depthwise and pointwise convolutions, specifically tailored to strike an optimal balance between processing efficiency and representational capacity. Within the MLFD pipeline, the data flow implements an efficient cross-branch feature aggregation strategy. Given the input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} \in \mathbb {R}^{H \times W \times C_{\text {in}}}$$\end{document}, the initial \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 1$$\end{document} convolution generates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_0$$\end{document}. Subsequent branches \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y}_i$$\end{document} integrate the features from preceding branches in a cascaded manner, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {DBUv2}$$\end{document} is the core processing unit. The final aggregated feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{\text {out}}$$\end{document} is obtained by concatenating the results from all branches \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y}_{i}$$\end{document} and the initial feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_0$$\end{document}, followed by a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 1$$\end{document} fusion convolution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{W}_{\text {fus}}$$\end{document} and the SiLU activation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma (\cdot )$$\end{document}:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{X}_{\text {out}} = \sigma \left( \text {BN} \left( \textbf{W}_{\text {fus}} * \left[ \textbf{X}_{0} ; \textbf{Y}_{1} ; \ldots ; \textbf{Y}_{n} \right] \right) \right) \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y}_{i} = \text {DBUv2}(\textbf{X}_{i-1})$$\end{document} represents the i-th branch output with cascaded features, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\cdot ; \cdot ]$$\end{document} denotes channel concatenation. The DBUv2 block’s core efficiency stems from the Depthwise Separable Convolution (DWConv). For a kernel size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k \times k$$\end{document}, the computational cost ratio R between DWConv (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FLOPs}_{\text {DWConv}}$$\end{document}) and a Standard Convolution (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FLOPs}_{\text {StdConv}}$$\end{document}) of the same size and channel configuration is:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R = \frac{\text {FLOPs}_{\text {DWConv}}}{\text {FLOPs}_{\text {StdConv}}} \approx \frac{1}{C_{\text {out}}} + \frac{1}{k^2} \end{aligned}$$\end{document}This ratio \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R \ll 1$$\end{document} provides the rigorous mathematical justification for MLFD’s efficiency. Furthermore, the large kernel k in DBUv2 is chosen to drastically expand the Effective Receptive Field (ERF), enhancing the model’s theoretical representation capacity R. This architectural choice enhances feature reuse and information flow across the backbone. The MLFD module addresses critical challenges in real-world disease detection: efficiency and robust feature extraction. The parallel multi-branch structure increases representational flexibility, enabling the model to capture diverse lesion features (scales, orientations) present in early-stage disease. Crucially, the integration of large convolutional kernels allows the model to access significantly extended spatial context. This extended receptive field is vital for recognizing subtle lesions and integrating surrounding environmental information (e.g., leaf texture), which enhances accuracy in distinguishing disease from background noise. Simultaneously, the pervasive use of depthwise separable convolutions in the DBUv2 units ensures a substantial reduction in both parameter count and computational complexity, directly fulfilling the need for a high-efficiency backbone suitable for low-power or embedded hardware. The final cascaded ELAN-style feature aggregation strategy ensures optimal information flow and feature reuse, making MLFD an effective module for improving both accuracy and efficiency in real-time agricultural disease detection systems. MDTA: Multi-scale dilated transformer attention module Fig. 6Detailed Architecture of the Multi-scale Dilated Transformer Attention (MDTA) Module. This figure presents the three core components of the MDTA module. Subfigure (a) illustrates the Channel Group Attention, which utilizes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} dilated depthwise convolutions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Dconv}$$\end{document}) to extract features, grouping channels (by factor g) to compute Query (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Q}$$\end{document}), Key (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {K}$$\end{document}), and Value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {V}$$\end{document}) matrices, reducing computational complexity. Subfigure (b) shows the Dilated FFN (Feed-Forward Network), composed of two \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} dilated convolutions separated by a GELU activation function. Subfigure (c) details the Window Multihead Attention, which uses \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 1$$\end{document} convolutions and a window size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M^2 \times M^2$$\end{document}) to apply attention across spatial windows, enhancing local feature focus while maintaining computational efficiency. Detailed Architecture of the Multi-scale Dilated Transformer Attention (MDTA) Module. This figure presents the three core components of the MDTA module. Subfigure (a) illustrates the Channel Group Attention, which utilizes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} dilated depthwise convolutions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Dconv}$$\end{document}) to extract features, grouping channels (by factor g) to compute Query (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Q}$$\end{document}), Key (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {K}$$\end{document}), and Value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {V}$$\end{document}) matrices, reducing computational complexity. Subfigure (b) shows the Dilated FFN (Feed-Forward Network), composed of two \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} dilated convolutions separated by a GELU activation function. Subfigure (c) details the Window Multihead Attention, which uses \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times 1$$\end{document} convolutions and a window size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M^2 \times M^2$$\end{document}) to apply attention across spatial windows, enhancing local feature focus while maintaining computational efficiency. Accurate rice leaf disease detection requires robust feature representation that can withstand severe environmental variations and effectively link localized lesions to their global context. The baseline C2PSA attention module, relying primarily on local convolutional operations, is deficient in capturing long-range dependencies and global spatial context. To overcome this limitation, we propose the MDTA module. MDTA is designed to enhance feature robustness by integrating a Transformer-inspired attention mechanism with multi-scale dilated convolutions. The core architectural choice is the strategic use of dilation to achieve an exponentially expanded receptive field without introducing extra parameters, enabling the capture of non-local relationships essential for complex scene understanding and improved contextual awareness across feature scales. The main network structure of MDTA is illustrated in Fig. 6. The MDTA module integrates a distinct two-phase attention mechanism for comprehensive feature map refinement, consisting of the Channel Group Attention (CGA) and Window Multi-head Attention (WMA) components. These stages operate in sequence to augment mutually supportive information. The CGA mechanism, taking cues from efficient group-based interaction methodologies, prioritizes strengthening inter-channel relationships through the factorization of the input tensor into Query (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}$$\end{document}), Key (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{K}$$\end{document}), and Value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V}$$\end{document}) representations. This channel-based weighting is applied post-LayerNormalization–a step vital for training stability–to enable the network to prioritize the most diagnostically significant feature channels. Following this, the WMA module executes localized attention calculation within designated, non-overlapping spatial windows. This structural parallel, resembling the window attention utilized in the Swin-Transformer, is essential for efficiently extracting fine-grained spatial relationships and localized structure patterns while preserving computational parsimony. The foundational concept for the CGA design originates from the efficient group interaction methodology presented in Restormer40. The process initiates by using a convolutional layer to partition the input tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}$$\end{document} into its query (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}$$\end{document}), key (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{K}$$\end{document}), and value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V}$$\end{document}) components. The resulting attention map is derived by first conducting the scaled dot-product attention operation between the queries and keys, and then performing a weighted summation over the values. Crucially, the full input tensor undergoes normalization using LayerNorm prior to this, a step that stabilizes the feature distribution’s mean and variance, thereby accelerating the model’s convergence during training. Defining the input tensor as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} \in \mathbb {R}^{B \times C \times H \times W}$$\end{document}, where B, C, H, and W specify the batch size, number of channels, spatial height, and width, respectively. The attention weights for the channel dimension, denoted \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}_{\text {channel}}$$\end{document}, are formally calculated as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{A}_{\text {channel}} = \text {softmax} \left( \frac{\textbf{Q} \textbf{K}^{\top }}{\sqrt{d_k}} \right) \end{aligned}$$\end{document}Here, the matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{K}$$\end{document} are represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}, \textbf{K} \in \mathbb {R}^{C \times d_k}$$\end{document}, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_k$$\end{document} signifying the key vector dimension. The derived result \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}_{\text {channel}}$$\end{document} is subsequently applied as weights to the value matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V}$$\end{document}, producing a refined feature map with strengthened channel-wise relationships. Following the channel refinement, the resulting tensor proceeds to the WMA module, which executes attention exclusively on local regions defined by fixed-size windows. This architectural choice bears conceptual resemblance to the window-based self-attention mechanism established in the Swin-Transformer framework41. The input feature map is systematically divided into distinct, non-overlapping spatial windows. Inside each window, the necessary queries, keys, and values are derived using pointwise (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 1$$\end{document}) convolutions. The spatial attention matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}_{\text {spatial}}^{W}$$\end{document} for a given window W is then computed according to:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{A}_{\text {spatial}}^{W} = \text {softmax} \left( \frac{\textbf{Q}^{W} (\textbf{K}^{W})^{\top }}{\sqrt{d_q}} \right) \end{aligned}$$\end{document}Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}^{W}, \textbf{K}^{W}$$\end{document} denote the queries and keys within window W, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_q$$\end{document} is the dimensionality of the query vectors. Following window-wise attention, the individually processed patches are seamlessly reassembled through concatenation to reconstruct the full-resolution feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{\text {att}}$$\end{document}. Following the sequential application of CGA and WMA, the MDTA module employs a Feed-Forward Network (FFN) to further process the attended features. This FFN is a critical architectural choice, as it is constructed using dilated convolutions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Conv}_{r}(\cdot )$$\end{document} (also known as atrous convolutions) rather than standard convolutions. This design deepens the feature abstraction while actively maintaining and leveraging the large, non-local receptive field established by the Transformer-style attention. Specifically, the operation of the FFN block, which incorporates a skip connection, can be formally described as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{X}_{\text {FFN}} = \textbf{X}_{\text {att}} + \sigma \left( \text {Conv}_{r_2}(\textbf{W}_{1} * \sigma (\text {Conv}_{r_1}(\textbf{W}_{0} * \textbf{X}_{\text {att}}))) \right) \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_2$$\end{document} are the dilation rates used in the FFN’s two convolutional layers. The inclusion of the dilation rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r> 1$$\end{document} ensures that the FFN actively maintains and leverages the large, non-local receptive field established by the preceding attention mechanisms. This structural coupling of attention and dilated FFN enhances the overall robustness and feature discrimination capability of the MDTA module. The MDTA module’s architecture provides a significant advantage in capturing the fine-grained and heterogeneous characteristics of rice leaf lesions. The combination of CGA and WMA creates a powerful dual-attention mechanism, allowing the network to selectively focus on crucial diagnostic features across both the channel and spatial dimensions. Crucially, the integration of dilated convolutions ensures the model possesses the requisite global contextual awareness, improving its capability to handle the spatial heterogeneity and occlusions prevalent in field environments. This ability to integrate long-range context and local fine-grained details is the key technical contributor to the model’s superior robustness and accuracy compared to the baseline C2PSA. Lo-Head: Lightweight Detection Head Fig. 7Detailed Architecture and Output Flow of the Lightweight Detection Head (Lo-Head). The figure illustrates the architecture of the Lo-Head. Feature Maps from the Neck enter the Stem, which preprocesses them via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolutions. The processed features are then split into parallel branches (Cv2 and Cv3) for regression (Bbox feature) and classification (cls feature) respectively, across different detection layers (Map1, Map2, etc.). The features are concatenated to form the final outputs (feature-0, feature-1, etc.). The outputs are subsequently processed for loss calculation, including the Distribution Focal Loss (DFL) for bounding box regression and the Classification Loss. Simultaneously, the outputs are used to generate Decoded Bboxes and Anchors for final detection results. Detailed Architecture and Output Flow of the Lightweight Detection Head (Lo-Head). The figure illustrates the architecture of the Lo-Head. Feature Maps from the Neck enter the Stem, which preprocesses them via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolutions. The processed features are then split into parallel branches (Cv2 and Cv3) for regression (Bbox feature) and classification (cls feature) respectively, across different detection layers (Map1, Map2, etc.). The features are concatenated to form the final outputs (feature-0, feature-1, etc.). The outputs are subsequently processed for loss calculation, including the Distribution Focal Loss (DFL) for bounding box regression and the Classification Loss. Simultaneously, the outputs are used to generate Decoded Bboxes and Anchors for final detection results. The Lo-Head exhibits notable performance advantages for rice leaf disease detection, particularly excelling in the minimization of computational complexity and the enhancement of model adaptability. This provides a pragmatic alternative for achieving efficient and accurate disease identification. Its network topology is visualized in Fig. 7: The Stem unit within the Lo-Head detection structure incorporates two sequential \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} group convolutional layers, which effectively curb both the parameter count and the overall computational load. Group convolution operates by segmenting the input channels into several distinct groups, with each group undergoing independent convolution. The resulting feature maps are then aggregated via concatenation along the channel dimension to construct the complete feature representation. This architecture not only reduces computational intensity but also retains essential spatial and semantic fidelity within the features. Consequently, this module is ideally suited for deployment on resource-constrained embedded platforms, facilitating real-time operational capacity in agricultural settings. Lo-Head ensures effective feature integration through the deployment of two specialized branches: the cv2 module, which is dedicated to extracting bounding box regression features, and the cv3 module, responsible for generating class probability features. During the feature combination stage, the outputs from these two branches are concatenated channel-wise, allowing the network to synthesize spatial and semantic information into a unified representation. This holistic fusion process substantially improves the model’s decision-making capacity when identifying lesion regions, especially under challenging field conditions. By capitalizing on the complementary nature of localization and classification features, the model achieves superior detection performance for rice leaf diseases. The Distribution Focal Loss (DFL) module in Lo-Head can be conditionally activated depending on the value of the reg_max parameter. This mechanism is critical for achieving sub-pixel level localization precision, which is vital for accurately bounding small lesions with ambiguous edges. When reg_max > 1, the DFL mechanism models the discrete probability distribution P(z) of bounding box offsets z. During inference, the predicted distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}$$\end{document} is converted into continuous offsets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_{\text {out}}$$\end{document} using the expectation operator \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {E}[\cdot ]$$\end{document}, scaled by the feature map’s stride. The computation is formally expressed as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{x}_{\text {out}} = \mathbb {E}[\textbf{b}] \cdot \textrm{stride} = \left( \sum _{i=0}^{n} z_i \cdot P(z_i) \right) \cdot \textrm{stride} \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_{\text {out}}$$\end{document} represents the final bounding box coordinates, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}$$\end{document} represents the predicted distribution of offsets, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{z_i\}$$\end{document} are the predefined discrete offset values with predicted probabilities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P(z_i)$$\end{document}. This expectation-based formulation rigorously justifies the model’s capacity to capture fine-grained variations in object localization. This methodology enables the model to more effectively capture minute variations in both the size and spatial positioning of lesions, thereby enhancing the localization accuracy for subtle or morphologically complex disease regions. If the condition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\texttt {reg\_max} \le 1$$\end{document} is met, the DFL module is intentionally bypassed to eliminate redundant computation, which boosts the model’s overall adaptability and operational efficiency. During the inference phase, the Lo-Head decodes the output from the DFL module by applying a geometric transformation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi$$\end{document}. This function translates the distribution-based offset representation back into concrete bounding box coordinates. The entire decoding process is mathematically defined based on the input feature tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{\text {cat}}$$\end{document}: The first step involves splitting the fused output feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{\text {cat}}$$\end{document} along the channel dimension C. Given the total localization dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d = \text {reg}_{\text {max}} \times 4$$\end{document}, the features are decoupled into the localization distribution tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{B}$$\end{document} and the classification logits tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{C}$$\end{document}:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{B} = \textbf{X}_{\text {cat}}[:, :, 0:d], \quad \textbf{C} = \textbf{X}_{\text {cat}}[:, :, d:C_{\text {total}}] \end{aligned}$$\end{document}Next, the final geometric bounding box tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{D}_{\text {box}}$$\end{document} is derived by applying the decoding transformation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi (\cdot )$$\end{document} to the localization features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{B}$$\end{document} and the anchor priors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}$$\end{document}:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{D}_{\text {box}} = \Phi (\textbf{B}, \textbf{A}) \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi \equiv \text {dist2bbox}(\text {DFL}(\cdot ), \cdot )$$\end{document}. Finally, the complete prediction output tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y}$$\end{document} is formed by concatenating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{D}_{\text {box}}$$\end{document} and the sigmoid-activated class probabilities, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma (\textbf{C}) \in (0, 1)$$\end{document}:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{Y} = \left[ \textbf{D}_{\text {box}} ; \sigma (\textbf{C}) \right] \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[\cdot ; \cdot ]$$\end{document} denotes concatenation along the channel dimension, yielding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y} \in \mathbb {R}^{H \times W \times (4+N_c)}$$\end{document}. In Equation 7, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}_{\text {cat}}$$\end{document} is the spliced feature map, and the operation defines the tensor slices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{B}$$\end{document} (bounding box features) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{C}$$\end{document} (category probability features) based on the dimension boundary d. In Equation 8, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi (\textbf{B}, \textbf{A})$$\end{document} represents the non-linear decoding operation that converts the DFL-processed offset distributions into specific bounding box coordinates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{D}_{\text {box}}$$\end{document}, utilizing the preset anchor parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{A}$$\end{document} and feature map strides. In Equation 9, the sigmoid activation function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma (\cdot )$$\end{document} converts the category logits into robust probability values, which are then concatenated with the decoded coordinates to obtain the final output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y}$$\end{document}, providing accurate localization and classification results for disease detection. The detection head implements a capability for the on-the-fly generation of anchor points and strides, scaling them relative to the input feature map’s geometry during the training cycle. This dynamic mechanism guarantees the model’s inherent ability to accommodate input images of disparate dimensions. Such adaptive scaling permits the framework to sustain reliable detection performance across the diverse leaf morphologies and sizes typical of various rice growth stages and cultivars, thereby enhancing the model’s overall generalization and operational flexibility. Pertaining to the small disease regions commonly found in rice leaf detection, the Lo-Head module is proficient at capturing the minute, detailed features of these targets, a capacity enabled by its lightweight yet efficient feature extraction coupled with its robust feature fusion capability. In conclusion, the Lo-Head detection head demonstrates considerable advantages in the rice leaf disease detection task. This superiority is rooted in its efficient, low-overhead feature extraction, powerful feature fusion capabilities, the adaptive deployment of the DFL mechanism, precise bounding box decoding and probability calculation, and outstanding efficacy in small-target detection. The resulting architecture not only achieves higher detection precision but also reduces the requisite computational resource load. This combined benefit ensures the model is optimally adapted for the complex operational conditions and diversified requirements encountered in practical applications, offering significant technological backing for the timely identification and precise intervention for rice leaf diseases. Experimental setup and evaluation metrics Hardware and software platform All experimental trials were executed on a dedicated computing workstation operating under the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Ubuntu 22.04.3 LTS}$$\end{document} (64-bit) environment. The underlying hardware platform was provisioned with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Intel}^{\circledR }$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Core}^{\textrm{TM}} \text {i9-14900K}$$\end{document} processor clocked at 4.20 GHz, complemented by 80 GB of system memory, and featuring an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {NVIDIA GeForce RTX 4090}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GPU}$$\end{document} equipped with 24 GB of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {VRAM}$$\end{document}. The utilized software infrastructure comprised \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {NVIDIA}$$\end{document} driver version 550.100, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {PyTorch 2.6.0}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {CUDA 12.4}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Python 3.10.8}$$\end{document}. Hyper-parameter configuration and training protocol Hyper-parameter selection strategy To determine the optimal configuration that maximizes performance while maintaining computational efficiency, we employed a systematic Grid Search strategy. This approach involved defining a search space for critical hyper-parameters, including \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lr0}$$\end{document} (initial learning rate), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {weight\_decay}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {batch}$$\end{document} size. The final hyper-parameter values, as detailed in Table 2, were selected based on achieving the highest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} score after preliminary 50-epoch training runs, ensuring robust and generalizable performance across the dataset. The model was trained for a total of 300 epochs using the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Stochastic Gradient Descent}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {SGD}$$\end{document}) optimizer, which utilized a 0.937 momentum parameter and a weight decay of 0.0005. The complete configuration of hyper-parameters, essential for ensuring the reproducibility of our work, is detailed in Table 2.Table 2Comprehensive Hyper-parameter Configuration for Model Training.CategoryParameterValueInput ConfigurationImage Resolution (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {imgsz}$$\end{document})\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$640 \times 640$$\end{document}Training Batch Size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {batch}$$\end{document})32Inference Batch Size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {batch}$$\end{document})32Optimizer (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {SGD}$$\end{document})Optimizer Type\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {SGD}$$\end{document}Momentum0.937Weight Decay0.0005Learning Rate ScheduleTotal Epochs300Initial Learning Rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lr0}$$\end{document})0.01Final Learning Rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {LR}_{\text {final}}$$\end{document})\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-4}$$\end{document}(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lr0} \times \text {lrf}=0.01 \times 0.01$$\end{document})LR Scheduler TypeLinear Decay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {cos\_lr}: \text {False}$$\end{document})Warmup Epochs (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {warmup\_epochs}$$\end{document})3.0Warmup Momentum (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {warmup\_momentum}$$\end{document})0.8Warmup Bias LR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {warmup\_bias\_lr}$$\end{document})0.1Augmentation StrategyMosaic Probability (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mosaic}$$\end{document})1.0Close Mosaic Epochs (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {close\_mosaic}$$\end{document})10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {hsv\_s}$$\end{document}/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {hsv\_v}$$\end{document}0.7/0.4Training StrategyAutomatic Mixed Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {amp}$$\end{document})TrueDevice’0’ (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GPU}$$\end{document}) Table 2Comprehensive Hyper-parameter Configuration for Model Training.CategoryParameterValueInput ConfigurationImage Resolution (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {imgsz}$$\end{document})\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$640 \times 640$$\end{document}Training Batch Size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {batch}$$\end{document})32Inference Batch Size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {batch}$$\end{document})32Optimizer (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {SGD}$$\end{document})Optimizer Type\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {SGD}$$\end{document}Momentum0.937Weight Decay0.0005Learning Rate ScheduleTotal Epochs300Initial Learning Rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lr0}$$\end{document})0.01Final Learning Rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {LR}_{\text {final}}$$\end{document})\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-4}$$\end{document}(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lr0} \times \text {lrf}=0.01 \times 0.01$$\end{document})LR Scheduler TypeLinear Decay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {cos\_lr}: \text {False}$$\end{document})Warmup Epochs (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {warmup\_epochs}$$\end{document})3.0Warmup Momentum (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {warmup\_momentum}$$\end{document})0.8Warmup Bias LR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {warmup\_bias\_lr}$$\end{document})0.1Augmentation StrategyMosaic Probability (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mosaic}$$\end{document})1.0Close Mosaic Epochs (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {close\_mosaic}$$\end{document})10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {hsv\_s}$$\end{document}/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {hsv\_v}$$\end{document}0.7/0.4Training StrategyAutomatic Mixed Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {amp}$$\end{document})TrueDevice’0’ (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GPU}$$\end{document}) Comprehensive Hyper-parameter Configuration for Model Training. The learning rate schedule employed a Linear Decay starting with an initial learning rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lr0}$$\end{document}) of 0.01 and reducing it to a final value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-4}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {lrf}=0.01 \times \text {lr0}$$\end{document}). A warm-up strategy was implemented to stabilize initial convergence. The learning rate warm-up lasted for the first 3.0 epochs, during which the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {SGD}$$\end{document} momentum started at 0.8. The bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {LR}$$\end{document} began at a higher value of 0.1 to accelerate bias term convergence. These strategies were essential for robust training and effective convergence. Cross-validation strategy To ensure the statistical robustness and generalization capability of our final reported results, we employed a 5-Fold Cross-Validation strategy. The entire Rice Leaf Disease Detection Dataset was initially and comprehensively partitioned into five non-overlapping subsets (folds) of equal size. This strict partitioning guarantees that every data point is used exactly once for validation. The model was subsequently trained and evaluated five times, with each of the five folds serving exactly once as the validation set while the remaining four folds were collectively used for training (Fig. 8). The final performance metrics for the proposed model (as reported in Table 5, Figs.9 and 10) are presented as the mean and standard deviation across the results obtained from the five independent validation runs. This rigorous methodology confirms that the reported performance is not merely an artifact of a single, fortuitous train-test split, significantly boosting the credibility of our findings. Unified training and comparison protocol To ensure a fair and rigorous comparison, we adopted a unified training protocol for all baseline models and the proposed framework. All compared architectures, including \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv5n}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv8n}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv10n}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv12}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv13}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Faster R-CNN}$$\end{document}, were trained using the exact same hyper-parameter configuration detailed in Table 2. This approach eliminates performance bias stemming from disparate training strategies. Specifically, the following consistency measures were strictly enforced: Implementation Source: All \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLO}$$\end{document} models were implemented and trained using a standardized framework based on the Ultralytics official repository, ensuring a consistent codebase and training pipeline for fair comparison. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Faster R-CNN}$$\end{document} baseline was implemented using the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MMEngine}$$\end{document} framework with equivalent configuration settings.Training from Scratch: To isolate the performance impact of our architectural improvements, all models, including the baselines, were initialized randomly and trained exclusively on our Rice Leaf Disease Detection Dataset from scratch. No external pre-trained weights (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {ImageNet}$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {COCO}$$\end{document}) were utilized. They were all trained for the full 300 epochs.FLOPs and FPS Resolution: The computational complexity (GFLOPs) and inference speed (FPS) for all models were consistently evaluated using an exact input image resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$640 \times 640$$\end{document}, ensuring that all reported hardware performance metrics are directly comparable.Training Time and Convergence Analysis: We meticulously recorded the total training time and monitored the convergence speed (epochs to reach 90% of final mAP) for all models. This data, presented in the Discussion (Sect. "Analysis of core technical contributions and performance" ), allows for a direct comparison of the training efficiency and scalability of our proposed architecture against established baselines.These measures confirm that the observed performance gains of our proposed model are attributable solely to the architectural improvements (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MLFD}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MDTA}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Lo-Head}$$\end{document}), rather than reliance on pre-trained knowledge or differences in training setup. Implementation Source: All \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLO}$$\end{document} models were implemented and trained using a standardized framework based on the Ultralytics official repository, ensuring a consistent codebase and training pipeline for fair comparison. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Faster R-CNN}$$\end{document} baseline was implemented using the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MMEngine}$$\end{document} framework with equivalent configuration settings. Implementation Source: All \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLO}$$\end{document} models were implemented and trained using a standardized framework based on the Ultralytics official repository, ensuring a consistent codebase and training pipeline for fair comparison. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Faster R-CNN}$$\end{document} baseline was implemented using the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MMEngine}$$\end{document} framework with equivalent configuration settings. Training from Scratch: To isolate the performance impact of our architectural improvements, all models, including the baselines, were initialized randomly and trained exclusively on our Rice Leaf Disease Detection Dataset from scratch. No external pre-trained weights (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {ImageNet}$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {COCO}$$\end{document}) were utilized. They were all trained for the full 300 epochs. Training from Scratch: To isolate the performance impact of our architectural improvements, all models, including the baselines, were initialized randomly and trained exclusively on our Rice Leaf Disease Detection Dataset from scratch. No external pre-trained weights (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {ImageNet}$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {COCO}$$\end{document}) were utilized. They were all trained for the full 300 epochs. FLOPs and FPS Resolution: The computational complexity (GFLOPs) and inference speed (FPS) for all models were consistently evaluated using an exact input image resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$640 \times 640$$\end{document}, ensuring that all reported hardware performance metrics are directly comparable. FLOPs and FPS Resolution: The computational complexity (GFLOPs) and inference speed (FPS) for all models were consistently evaluated using an exact input image resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$640 \times 640$$\end{document}, ensuring that all reported hardware performance metrics are directly comparable. Training Time and Convergence Analysis: We meticulously recorded the total training time and monitored the convergence speed (epochs to reach 90% of final mAP) for all models. This data, presented in the Discussion (Sect. "Analysis of core technical contributions and performance" ), allows for a direct comparison of the training efficiency and scalability of our proposed architecture against established baselines. Training Time and Convergence Analysis: We meticulously recorded the total training time and monitored the convergence speed (epochs to reach 90% of final mAP) for all models. This data, presented in the Discussion (Sect. "Analysis of core technical contributions and performance" ), allows for a direct comparison of the training efficiency and scalability of our proposed architecture against established baselines. Evaluation metrics The model’s performance was evaluated using standard object detection metrics, including \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GFLOPs}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPS}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Precision}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {P}$$\end{document}), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Recall}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {R}$$\end{document}), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {F1-Score}$$\end{document}, and mean Average Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document}). The definitions are as follows:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Precision} = \frac{TP}{TP + FP} \end{aligned}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Recall} = \frac{TP}{TP + FN} \end{aligned}$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {F1-Score} = 2 \times \frac{\text {Precision} \times \text {Recall}}{\text {Precision} + \text {Recall}} \end{aligned}$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {mAP} = \frac{\sum _{i=1}^{N} \int _{0}^{1} P(R_i) \, dR}{N} \end{aligned}$$\end{document}In this context, TP represents true positives (instances of disease regions correctly identified), FP signifies false positives (erroneous detections), and FN refers to false negatives (missed disease regions). N is defined as the total number of distinct categories. The Mean Average Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document}) metric is assessed using two standard configurations: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document}. The former is calculated using a singular, fixed Intersection over Union (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {IoU}$$\end{document}) threshold of 0.5. In contrast, the latter variant computes the average of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}$$\end{document} values across a range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {IoU}$$\end{document} thresholds, spanning from 0.5 to 0.95 in increments of 0.05. Each individual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}$$\end{document} (Average Precision) value is quantified as the area underneath its corresponding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Precision-Recall}$$\end{document} curve. Finally, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document} is obtained by taking the average of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}$$\end{document} scores across all detected categories. Experimental results Selection of baseline model To establish the efficacy of different \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11}$$\end{document} variants for rice leaf disease detection, five versions of the network models were trained under identical experimental conditions. The comparative performance metrics are presented in Table 3. Analysis of this table reveals that, in comparison to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11s}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11m}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11l}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11x}$$\end{document}, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document} model exhibits a lower Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {P}$$\end{document}) by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.87\%$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.85\%$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.26\%$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.99\%$$\end{document} respectively, along with a correspondingly lower Recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {R}$$\end{document}). Critically, however, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} performance gap between the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document} version and the four larger variants is less than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\%$$\end{document}. With a computational volume of 6.3 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GFLOPs}$$\end{document}, 2.58 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {M}$$\end{document} parameters, and a model size of 5.2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {M}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document} is significantly superior to the other four models in terms of resource efficiency. Considering this minimal compromise in detection accuracy (Precision, Recall, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document}) against the vast improvement in network efficiency, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document} was selected as the designated benchmark model for this study. A further justification for this choice is its inherent suitability for deployment on resource-limited edge devices and its practical relevance for real-time, on-site detection in natural field environments.Table 3Performance comparison of different YOLOv11 model variants on the new dataset.ModelFLOPs (G)Params (M)PrecisionRecallmAP@0.5Model Size (MB)yolov11-n6.32.580.82870.78560.84495.2yolov11-s21.39.410.84510.84020.849318.3yolov11-m67.720.030.85350.84630.851038.7yolov11-l86.625.280.84850.84900.852148.9yolov11-x194.456.830.85470.84450.8531109.1 Table 3Performance comparison of different YOLOv11 model variants on the new dataset.ModelFLOPs (G)Params (M)PrecisionRecallmAP@0.5Model Size (MB)yolov11-n6.32.580.82870.78560.84495.2yolov11-s21.39.410.84510.84020.849318.3yolov11-m67.720.030.85350.84630.851038.7yolov11-l86.625.280.84850.84900.852148.9yolov11-x194.456.830.85470.84450.8531109.1 Performance comparison of different YOLOv11 model variants on the new dataset. Ablation experiments To quantitatively ascertain the distinct and synergistic contributions of the Multi-Branch Large-Kernel Fusion Depthwise (MLFD), Multi-Scale Dilated Transformer Attention (MDTA), and Lo-Head modules within our proposed architecture, we executed a comprehensive series of ablation studies. The detailed outcomes from this experimental analysis are compiled in Table 4.Table 4Performance comparison of different model variants.MLFDMDTALo-HeadParams (M)FLOPs (G)FPSPrecisionRecallmAP@0.5mAP@0.5:0.95ModelSize (MB)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.586.3145.940.82870.78560.84490.60855.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.927.2104.830.82960.77510.84150.60516.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.897.5114.810.81680.78860.84360.61716.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.315.1153.630.82540.79870.85240.61594.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.937.2101.750.8330.77570.83880.60926.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.656121.120.82970.790.84470.61476\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.435.1113.950.83250.78790.84920.62375.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.666.3106.480.84390.78610.84980.62625.6 Table 4Performance comparison of different model variants.MLFDMDTALo-HeadParams (M)FLOPs (G)FPSPrecisionRecallmAP@0.5mAP@0.5:0.95ModelSize (MB)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.586.3145.940.82870.78560.84490.60855.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.927.2104.830.82960.77510.84150.60516.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.897.5114.810.81680.78860.84360.61716.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.315.1153.630.82540.79870.85240.61594.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}2.937.2101.750.8330.77570.83880.60926.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.656121.120.82970.790.84470.61476\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.435.1113.950.83250.78790.84920.62375.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}2.666.3106.480.84390.78610.84980.62625.6 Performance comparison of different model variants. Contribution of individual modules The baseline model establishes the performance benchmark, achieving a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6085}$$\end{document} with a lightweight model size of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.2 MB}$$\end{document}.MLFD Module: Introducing MLFD alone slightly improves feature representation capability. It achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6051}$$\end{document} and a precision of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8296}$$\end{document}. Compared to the baseline \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6085}$$\end{document}), this variant shows a minor drop (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {-0.0034}$$\end{document}), but its enhanced feature diversity for subtle lesions is noted, incurring an acceptable increase in FLOPs (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.3 G}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7.2 G}$$\end{document}). This validates the efficacy of the large-kernel depthwise structure in expanding the Effective Receptive Field (ERF).MDTA Module: MDTA achieves a significant \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6171}$$\end{document}, representing an improvement of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.86}$$\end{document} percentage points (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6171} - \text {0.6085}$$\end{document}). This validates the proposed multi-head attention mechanism’s superior ability to capture long-range dependencies and global contextual information, which is crucial for robust disease identification. This benefit comes with the largest computational overhead (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.7 MB}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7.5 GFLOPs}$$\end{document}), underscoring the trade-off inherent in attention-based designs.Lo-Head Module: The Lo-Head is primarily designed for efficiency. Its introduction substantially reduces the model size (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.2 MB}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {4.7 MB}$$\end{document}) and FLOPs (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.3 G}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.1 G}$$\end{document}). It achieves the highest FPS (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {153.63}$$\end{document}) among all single-module variants and significantly improves \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6159}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.74}$$\end{document} percentage points improvement), confirming its efficiency-first design goal while also boosting accuracy, making it highly suitable for resource-constrained edge deployment. MLFD Module: Introducing MLFD alone slightly improves feature representation capability. It achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6051}$$\end{document} and a precision of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8296}$$\end{document}. Compared to the baseline \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6085}$$\end{document}), this variant shows a minor drop (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {-0.0034}$$\end{document}), but its enhanced feature diversity for subtle lesions is noted, incurring an acceptable increase in FLOPs (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.3 G}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7.2 G}$$\end{document}). This validates the efficacy of the large-kernel depthwise structure in expanding the Effective Receptive Field (ERF). MLFD Module: Introducing MLFD alone slightly improves feature representation capability. It achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6051}$$\end{document} and a precision of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8296}$$\end{document}. Compared to the baseline \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6085}$$\end{document}), this variant shows a minor drop (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {-0.0034}$$\end{document}), but its enhanced feature diversity for subtle lesions is noted, incurring an acceptable increase in FLOPs (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.3 G}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7.2 G}$$\end{document}). This validates the efficacy of the large-kernel depthwise structure in expanding the Effective Receptive Field (ERF). MDTA Module: MDTA achieves a significant \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6171}$$\end{document}, representing an improvement of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.86}$$\end{document} percentage points (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6171} - \text {0.6085}$$\end{document}). This validates the proposed multi-head attention mechanism’s superior ability to capture long-range dependencies and global contextual information, which is crucial for robust disease identification. This benefit comes with the largest computational overhead (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.7 MB}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7.5 GFLOPs}$$\end{document}), underscoring the trade-off inherent in attention-based designs. MDTA Module: MDTA achieves a significant \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6171}$$\end{document}, representing an improvement of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.86}$$\end{document} percentage points (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6171} - \text {0.6085}$$\end{document}). This validates the proposed multi-head attention mechanism’s superior ability to capture long-range dependencies and global contextual information, which is crucial for robust disease identification. This benefit comes with the largest computational overhead (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.7 MB}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7.5 GFLOPs}$$\end{document}), underscoring the trade-off inherent in attention-based designs. Lo-Head Module: The Lo-Head is primarily designed for efficiency. Its introduction substantially reduces the model size (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.2 MB}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {4.7 MB}$$\end{document}) and FLOPs (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.3 G}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.1 G}$$\end{document}). It achieves the highest FPS (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {153.63}$$\end{document}) among all single-module variants and significantly improves \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6159}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.74}$$\end{document} percentage points improvement), confirming its efficiency-first design goal while also boosting accuracy, making it highly suitable for resource-constrained edge deployment. Lo-Head Module: The Lo-Head is primarily designed for efficiency. Its introduction substantially reduces the model size (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.2 MB}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {4.7 MB}$$\end{document}) and FLOPs (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.3 G}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.1 G}$$\end{document}). It achieves the highest FPS (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {153.63}$$\end{document}) among all single-module variants and significantly improves \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6159}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.74}$$\end{document} percentage points improvement), confirming its efficiency-first design goal while also boosting accuracy, making it highly suitable for resource-constrained edge deployment. Synergistic effects of module combinations Analyzing combinations of modules reveals their synergistic effects on performance and efficiency:MLFD + MDTA: This combination achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6092}$$\end{document}. While the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} is comparable to the baseline, the high Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8330}$$\end{document}) and low Recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.7757}$$\end{document}) suggest a highly confident, albeit slightly more conservative, detection capability, demonstrating the combination of local and global feature fusion.MDTA + Lo-Head: This configuration achieves a balanced result with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6147}$$\end{document}. Compared to the MDTA-only variant, the model size is reduced significantly (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.7 MB}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.0 MB}$$\end{document}), and FPS improves (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {114.81}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {121.12}$$\end{document}). This proves Lo-Head’s effectiveness in lightweighting the network without major performance degradation.MLFD + Lo-Head: This combination yields a strong \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6237}$$\end{document}, an improvement of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.52}$$\end{document} percentage points over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6237} - \text {0.6085}$$\end{document}). The model size remains highly efficient at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.4 MB}$$\end{document}, demonstrating that Lo-Head successfully compensates for the complexity introduced by MLFD while capitalizing on its feature enhancement capabilities. MLFD + MDTA: This combination achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6092}$$\end{document}. While the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} is comparable to the baseline, the high Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8330}$$\end{document}) and low Recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.7757}$$\end{document}) suggest a highly confident, albeit slightly more conservative, detection capability, demonstrating the combination of local and global feature fusion. MLFD + MDTA: This combination achieves a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6092}$$\end{document}. While the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} is comparable to the baseline, the high Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8330}$$\end{document}) and low Recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.7757}$$\end{document}) suggest a highly confident, albeit slightly more conservative, detection capability, demonstrating the combination of local and global feature fusion. MDTA + Lo-Head: This configuration achieves a balanced result with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6147}$$\end{document}. Compared to the MDTA-only variant, the model size is reduced significantly (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.7 MB}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.0 MB}$$\end{document}), and FPS improves (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {114.81}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {121.12}$$\end{document}). This proves Lo-Head’s effectiveness in lightweighting the network without major performance degradation. MDTA + Lo-Head: This configuration achieves a balanced result with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6147}$$\end{document}. Compared to the MDTA-only variant, the model size is reduced significantly (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.7 MB}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {6.0 MB}$$\end{document}), and FPS improves (from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {114.81}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {121.12}$$\end{document}). This proves Lo-Head’s effectiveness in lightweighting the network without major performance degradation. MLFD + Lo-Head: This combination yields a strong \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6237}$$\end{document}, an improvement of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.52}$$\end{document} percentage points over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6237} - \text {0.6085}$$\end{document}). The model size remains highly efficient at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.4 MB}$$\end{document}, demonstrating that Lo-Head successfully compensates for the complexity introduced by MLFD while capitalizing on its feature enhancement capabilities. MLFD + Lo-Head: This combination yields a strong \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6237}$$\end{document}, an improvement of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.52}$$\end{document} percentage points over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6237} - \text {0.6085}$$\end{document}). The model size remains highly efficient at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.4 MB}$$\end{document}, demonstrating that Lo-Head successfully compensates for the complexity introduced by MLFD while capitalizing on its feature enhancement capabilities. Proposed framework and overall performance Integrating all three proposed modules results in the final proposed framework. This integrated network successfully achieves the design goal of high accuracy with maintained efficiency:It achieves the highest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262}$$\end{document}, which is a substantial \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.77}$$\end{document} percentage point gain over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262} - \text {0.6085}$$\end{document}).The model size is highly optimized at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB}$$\end{document}, adding only \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.4 MB}$$\end{document} compared to the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB} - \text {5.2 MB}$$\end{document}), but delivering a substantial performance gain.The final model maintains a competitive FPS of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {106.48}$$\end{document}, confirming the successful implementation of MLFD and MDTA’s feature enhancement while minimizing computational cost via the Lo-Head module.The integrated framework thus shows significant improvement in detection accuracy, robustness, and efficiency for rice leaf disease detection. Furthermore, a detailed class-wise performance analysis (see supplementary Fig. 8 and the confusion matrix (see supplementary Fig. 12 provide granular evidence that each module reduces false positives and false negatives for specific disease types, confirming the modules’ theoretical contributions.Fig. 8Class-wise performance comparison for all ablation models. It achieves the highest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262}$$\end{document}, which is a substantial \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.77}$$\end{document} percentage point gain over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262} - \text {0.6085}$$\end{document}). It achieves the highest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262}$$\end{document}, which is a substantial \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.77}$$\end{document} percentage point gain over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262} - \text {0.6085}$$\end{document}). The model size is highly optimized at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB}$$\end{document}, adding only \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.4 MB}$$\end{document} compared to the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB} - \text {5.2 MB}$$\end{document}), but delivering a substantial performance gain. The model size is highly optimized at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB}$$\end{document}, adding only \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.4 MB}$$\end{document} compared to the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB} - \text {5.2 MB}$$\end{document}), but delivering a substantial performance gain. The final model maintains a competitive FPS of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {106.48}$$\end{document}, confirming the successful implementation of MLFD and MDTA’s feature enhancement while minimizing computational cost via the Lo-Head module. The final model maintains a competitive FPS of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {106.48}$$\end{document}, confirming the successful implementation of MLFD and MDTA’s feature enhancement while minimizing computational cost via the Lo-Head module. Class-wise performance comparison for all ablation models. K-fold cross-validation for robustness analysis Fig. 9The mAP@0.5:0.95 scores of the proposed framework across the five folds of the cross-validation. This experiment was conducted on the full dataset without data augmentation to assess the model’s intrinsic robustness and stability against data partitioning. The red dashed line represents the mean mAP@0.5:0.95 (0.6173) over the five runs, demonstrating the consistent performance of the integrated MLFD-MDTA-Lo-Head architecture. Fig. 10The mAP@0.5:0.95 scores of the proposed framework across the five folds of the cross-validation. This experiment was conducted on the full dataset without data augmentation to assess the model’s intrinsic robustness and stability against data partitioning. The red dashed line represents the mean mAP@0.5:0.95 (0.6173) over the five runs, demonstrating the consistent performance of the integrated MLFD-MDTA-Lo-Head architecture. Table 5The result of training the final model on a 5-fold cross-validation (No Augmentation). Data fold Number of instances P R F1-score mAP@0.5 mAP@0.5:0.95 Fold_1133970.84050.75130.79350.82560.6094Fold_2133210.84360.79810.82020.84130.6358Fold_3133290.85240.74890.79740.81240.6087Fold_4132320.85010.80330.82590.84590.6277Fold_5132970.83740.76970.80590.86560.6049Mean (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu$$\end{document})13315 0.8448 0.7743 0.8079 0.8262 0.6173 Std (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document})59 0.0078 0.0237 0.0140 0.0153 0.0115 The mAP@0.5:0.95 scores of the proposed framework across the five folds of the cross-validation. This experiment was conducted on the full dataset without data augmentation to assess the model’s intrinsic robustness and stability against data partitioning. The red dashed line represents the mean mAP@0.5:0.95 (0.6173) over the five runs, demonstrating the consistent performance of the integrated MLFD-MDTA-Lo-Head architecture. The mAP@0.5:0.95 scores of the proposed framework across the five folds of the cross-validation. This experiment was conducted on the full dataset without data augmentation to assess the model’s intrinsic robustness and stability against data partitioning. The red dashed line represents the mean mAP@0.5:0.95 (0.6173) over the five runs, demonstrating the consistent performance of the integrated MLFD-MDTA-Lo-Head architecture. Table 5The result of training the final model on a 5-fold cross-validation (No Augmentation). Data fold Number of instances P R F1-score mAP@0.5 mAP@0.5:0.95 Fold_1133970.84050.75130.79350.82560.6094Fold_2133210.84360.79810.82020.84130.6358Fold_3133290.85240.74890.79740.81240.6087Fold_4132320.85010.80330.82590.84590.6277Fold_5132970.83740.76970.80590.86560.6049Mean (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu$$\end{document})13315 0.8448 0.7743 0.8079 0.8262 0.6173 Std (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document})59 0.0078 0.0237 0.0140 0.0153 0.0115 The result of training the final model on a 5-fold cross-validation (No Augmentation). To provide rigorous statistical validation for the high performance achieved by the proposed framework and to address the necessity for a scientific assessment of model stability, a 5-fold cross-validation (5-CV) experiment was conducted. This approach responds directly to the need for reporting variance and robust performance metrics, as is standard practice for high-accuracy models in complex detection tasks. The 5-CV procedure involved partitioning the entire dataset into five distinct folds, ensuring that each fold was used exactly once as the validation set while the remaining four folds were used for training. Crucially, this experiment was performed on the full dataset without the application of data augmentation, thus isolating the model’s intrinsic ability to generalize across different subsets of the original data. The primary objective was to report the mean (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu$$\end{document}) and standard deviation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}) of key performance indicators, providing a comprehensive measure of the model’s reliability and its sensitivity to the specific training/validation split. The detailed results of the 5-CV experiment are presented in Table 5, and the consistency of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} metric across all folds is visually depicted in Fig. 9. Mean Performance: The average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} across the five independent runs is 0.6173. This mean value is highly competitive and is closely aligned with the best single-run performance (0.6262), confirming that the reported peak accuracy is not a statistical anomaly but is representative of the model’s true capability. Model Stability: The stability of the proposed architecture is demonstrated by the low standard deviation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}) of the performance metrics. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} exhibits a deviation of only 0.0115, which is an excellent result for an object detection task on a complex disease dataset. Similarly, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Precision}$$\end{document} shows exceptional stability with a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} of 0.0078. This low variance is a powerful testament to the effectiveness of the integrated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MLFD}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MDTA}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Lo-Head}$$\end{document} modules, indicating that the final network structure is highly robust and consistently achieves strong generalization performance regardless of how the training data is partitioned. The results successfully validate the stability of the proposed framework, thereby strengthening the reliability and scientific rigor of our findings. Comparative analysis with state-of-the-art detectors In order to further validate the comprehensive effectiveness of the proposed framework for rice leaf disease detection, a comparison experiment was conducted against several established baselines and state-of-the-art detectors. These include the classical two-stage model Faster R-CNN, and popular lightweight single-stage models such as YOLOv5n, YOLOv8n, YOLOv10n, YOLOv12, YOLOv13, and the baseline YOLOv11n. All models were trained and evaluated in the same hardware and software environment, utilizing the same dataset to ensure fair comparison. The detailed experimental results are shown in Table 6.Furthermore, the training curves (Fig. 13) confirm the robustness of our architectural enhancements, showing that the proposed framework not only achieves a higher final accuracy but also maintains a competitive, stable convergence rate throughout the 300 epochs compared to the YOLOv11n baseline.Table 6Performance comparison of different models on the new dataset.ModelParams (M)FLOPs (G)FPSPrecisionRecallF1-ScoremAP@0.5mAP@0.75mAP@0.5:0.95Model Size (MB)YOLOv5n2.185.8179.820.81170.77600.79010.83240.59340.56734.5YOLOv8n2.696.8192.450.81520.77750.79230.83390.59830.57415.4YOLOv10n2.276.5147.650.79710.75430.76930.82710.58070.56145.5YOLOv11n2.586.3145.940.82870.78560.80210.84490.64390.60855.2YOLOv122.515.826.040.84110.78810.81240.84820.65610.62065.2YOLOv132.456.165.310.84240.77450.81110.84890.65890.62455.2Faster-RCNN41.3613739.290.72500.71210.71880.82450.63710.582131.6ours2.666.3106.480.84390.78610.81110.84980.65920.62625.6 Table 6Performance comparison of different models on the new dataset.ModelParams (M)FLOPs (G)FPSPrecisionRecallF1-ScoremAP@0.5mAP@0.75mAP@0.5:0.95Model Size (MB)YOLOv5n2.185.8179.820.81170.77600.79010.83240.59340.56734.5YOLOv8n2.696.8192.450.81520.77750.79230.83390.59830.57415.4YOLOv10n2.276.5147.650.79710.75430.76930.82710.58070.56145.5YOLOv11n2.586.3145.940.82870.78560.80210.84490.64390.60855.2YOLOv122.515.826.040.84110.78810.81240.84820.65610.62065.2YOLOv132.456.165.310.84240.77450.81110.84890.65890.62455.2Faster-RCNN41.3613739.290.72500.71210.71880.82450.63710.582131.6ours2.666.3106.480.84390.78610.81110.84980.65920.62625.6 Performance comparison of different models on the new dataset. Accuracy and efficiency comparison Fig. 11Precision-Recall (PR) Curves for Individual Disease Classes and All Classes. This figure illustrates the Precision-Recall (PR) curves for the detection of individual rice leaf disease classes and their aggregate performance. Each curve represents the trade-off between Precision and Recall for a specific class (Brown Spot (BS), Leaf Blast (LB), Bacterial Leaf Blight (BLB), Bacterial Leaf Streak (BLS)), with their respective average precision (AP) values indicated in the legend. The bold blue curve represents the overall PR curve for all classes, achieving a mean Average Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document}) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.850}$$\end{document} at an Intersection over Union (IoU) threshold of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.5}$$\end{document}. Higher curves and AP values signify superior detection performance. Fig. 12Normalized Confusion Matrix of Proposed Framework. The matrix displays the normalized classification results, confirming the model’s robustness by showing high true positive rates along the diagonal and minimized misclassifications between different disease types and the background. Precision-Recall (PR) Curves for Individual Disease Classes and All Classes. This figure illustrates the Precision-Recall (PR) curves for the detection of individual rice leaf disease classes and their aggregate performance. Each curve represents the trade-off between Precision and Recall for a specific class (Brown Spot (BS), Leaf Blast (LB), Bacterial Leaf Blight (BLB), Bacterial Leaf Streak (BLS)), with their respective average precision (AP) values indicated in the legend. The bold blue curve represents the overall PR curve for all classes, achieving a mean Average Precision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document}) of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.850}$$\end{document} at an Intersection over Union (IoU) threshold of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.5}$$\end{document}. Higher curves and AP values signify superior detection performance. Normalized Confusion Matrix of Proposed Framework. The matrix displays the normalized classification results, confirming the model’s robustness by showing high true positive rates along the diagonal and minimized misclassifications between different disease types and the background. The proposed framework exhibits markedly superior performance when benchmarked against the classical two-stage detector, Faster R-CNN. Specifically, while Faster R-CNN yields a Precision of 0.7250, a Recall of 0.7121, an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} of 0.8245, and an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of 0.5821, our architecture achieves substantial enhancements across the board. The proposed framework attained a Precision of 0.8439, a Recall of 0.7861, an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} of 0.8498, and registered the highest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} score at 0.6262. Furthermore, concerning inference speed, the proposed framework operates at a frame rate of 106.48 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPS}$$\end{document}. This speed dramatically outpaces the 39.29 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPS}$$\end{document} recorded by Faster R-CNN, despite the latter’s much larger model size (31.6 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MB}$$\end{document}). These results conclusively validate the proposed framework’s viability for real-time application in agricultural settings.Fig. 13Training Curves: Comparison of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} between the Proposed Framework and the YOLOv11n Baseline. The figure illustrates the convergence and final performance achieved by the proposed framework (red solid line) compared to the YOLOv11n baseline (blue dashed line) throughout the training process. The proposed model consistently achieves higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document} values, demonstrating faster convergence and superior final detection accuracy. Training Curves: Comparison of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} between the Proposed Framework and the YOLOv11n Baseline. The figure illustrates the convergence and final performance achieved by the proposed framework (red solid line) compared to the YOLOv11n baseline (blue dashed line) throughout the training process. The proposed model consistently achieves higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document} values, demonstrating faster convergence and superior final detection accuracy. When compared to state-of-the-art lightweight detectors, the proposed framework consistently outperforms them in detection accuracy. The baseline YOLOv11n achieves an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6085}$$\end{document}, which is 1.77 percentage points lower than our model’s \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262}$$\end{document}. While the YOLOv12 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6206}$$\end{document}) and YOLOv13 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6245}$$\end{document}) variants show competitive performance, they are also marginally surpassed by our proposed model. YOLOv8n achieves the highest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPS}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {192.45}$$\end{document}) but yields a lower \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.5741}$$\end{document}), confirming the successful balance of high accuracy and maintained efficiency in our integrated architecture. Visual and qualitative analysis The comprehensive performance evaluation is further fortified by a detailed visual and qualitative assessment of the model’s output and internal mechanisms. The Precision-Recall (PR) Curves (Fig. 11) clearly illustrate the classification stability, detailing the performance trade-off across individual disease classes. The high total area under the bold blue PR curve (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.850}$$\end{document} as noted in the figure legend) visually corroborates the impressive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8498}$$\end{document}. Furthermore, the Normalized Confusion Matrix (Fig. 12) provides essential granular evidence of the model’s classification robustness, specifically confirming its minimized rate of false positives originating from background interference (Fig. 13). Visualizing Internal Mechanisms To explicitly demonstrate the effectiveness of the proposed MLFD and MDTA modules in feature refinement, we present comparative feature activation heatmaps (Fig. 14). The visualization methodology adheres to strict standards: Activation Layer: The heatmaps capture the feature output from the Neck-Layer Feature Fusion Module (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {P3/P4/P5}$$\end{document} outputs of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPN/PAN}$$\end{document} structure), representing the integrated, multi-scale semantic features before final detection.Colormap and Intensity: A continuous colormap is used, where warm colors (yellow to red) denote regions of high feature activation critical for localization, and cool colors (blue to green) indicate low-contributing features.Normalization: The raw feature map values are individually normalized via Min-Max Scaling (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0 to 1}$$\end{document}) across all channels, ensuring visual fidelity and direct comparability between our model and the baseline.Qualitative evaluation based on these heatmaps and raw detection images further validates our architectural advantages. The heatmaps (Fig. 14) reveal that the proposed framework generates a significantly sharper and more tightly localized activation response over the actual lesion regions compared to the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document}), which often exhibits diffused or scattered attention. Concurrently, the visual comparison of detection images (Fig. 15) confirms this enhanced localization, showing that our framework produces fewer false positives and more accurately fitted bounding boxes around disease lesions than all competing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLO}$$\end{document} variants. These comprehensive visual results collectively underscore the model’s superior advantages in accuracy, robustness, and efficiency for practical smart agriculture applications.Fig. 14Comparison of heatmap images derived from the baseline and proposed framework. The figure illustrates the enhanced localization ability of the proposed framework, showing stronger and more accurately centered activation responses over the lesion regions compared to the baseline model (YOLOv11n).Fig. 15Visual comparison of detection results among various models (YOLOv5n, YOLOv8n, YOLOv10n, YOLOv11n... and Proposed Framework). This comparison demonstrates the superior detection accuracy of the proposed framework, exhibiting fewer false positives and better placement of bounding boxes around the disease lesions compared to the competing models. Activation Layer: The heatmaps capture the feature output from the Neck-Layer Feature Fusion Module (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {P3/P4/P5}$$\end{document} outputs of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPN/PAN}$$\end{document} structure), representing the integrated, multi-scale semantic features before final detection. Activation Layer: The heatmaps capture the feature output from the Neck-Layer Feature Fusion Module (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {P3/P4/P5}$$\end{document} outputs of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {FPN/PAN}$$\end{document} structure), representing the integrated, multi-scale semantic features before final detection. Colormap and Intensity: A continuous colormap is used, where warm colors (yellow to red) denote regions of high feature activation critical for localization, and cool colors (blue to green) indicate low-contributing features. Colormap and Intensity: A continuous colormap is used, where warm colors (yellow to red) denote regions of high feature activation critical for localization, and cool colors (blue to green) indicate low-contributing features. Normalization: The raw feature map values are individually normalized via Min-Max Scaling (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0 to 1}$$\end{document}) across all channels, ensuring visual fidelity and direct comparability between our model and the baseline. Normalization: The raw feature map values are individually normalized via Min-Max Scaling (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0 to 1}$$\end{document}) across all channels, ensuring visual fidelity and direct comparability between our model and the baseline. Comparison of heatmap images derived from the baseline and proposed framework. The figure illustrates the enhanced localization ability of the proposed framework, showing stronger and more accurately centered activation responses over the lesion regions compared to the baseline model (YOLOv11n). Visual comparison of detection results among various models (YOLOv5n, YOLOv8n, YOLOv10n, YOLOv11n... and Proposed Framework). This comparison demonstrates the superior detection accuracy of the proposed framework, exhibiting fewer false positives and better placement of bounding boxes around the disease lesions compared to the competing models. Robustness and error analysis Table 7COCO Evaluation Metrics Summary of the Proposed Model. Metric type IoU/Area Value Average Precision (AP) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}$$\end{document} @ [0.50 : 0.95]0.605 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{50}$$\end{document} @ [0.50]0.837 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document} @ [0.75]0.635 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}}$$\end{document} area < 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.291 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}}$$\end{document} 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document} < area < 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.672 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}}$$\end{document} area > 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.720Average Recall (AR) - Max Dets = 100 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}$$\end{document} @ [0.50 : 0.95]0.673 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {small}}$$\end{document} area < 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.329 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {medium}}$$\end{document} 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document} < area < 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.740 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {large}}$$\end{document} area > 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.774 Table 7COCO Evaluation Metrics Summary of the Proposed Model. Metric type IoU/Area Value Average Precision (AP) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}$$\end{document} @ [0.50 : 0.95]0.605 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{50}$$\end{document} @ [0.50]0.837 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document} @ [0.75]0.635 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}}$$\end{document} area < 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.291 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}}$$\end{document} 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document} < area < 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.672 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}}$$\end{document} area > 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.720Average Recall (AR) - Max Dets = 100 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}$$\end{document} @ [0.50 : 0.95]0.673 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {small}}$$\end{document} area < 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.329 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {medium}}$$\end{document} 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document} < area < 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.740 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {large}}$$\end{document} area > 96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document}0.774 COCO Evaluation Metrics Summary of the Proposed Model. To fully assess the model’s robustness and understand its failure modes, a detailed analysis based on COCO metrics and TIDE (Toolkit for Image Detection and Evaluation) was conducted. The detailed COCO metrics are presented in Table 7, while the TIDE error breakdown is visualized in Fig. 16.Scale Robustness (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}}$$\end{document}): The model demonstrates strong performance on medium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}} = 0.672$$\end{document}) and large (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}} = 0.720$$\end{document}) objects. However, performance significantly drops on small objects (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}} = 0.291$$\end{document}), suggesting that while the proposed MLFD and MDTA modules effectively enhance feature fusion, the detection of very subtle or distant lesions (small objects) remains the primary challenge. The corresponding recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {small}} = 0.329$$\end{document}) also confirms the difficulty in recovering these small instances.Localization Accuracy: The model achieves a competitive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{50}$$\end{document} of 0.837, indicating high accuracy at a loose Intersection over Union (IoU) threshold. However, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document} drops to 0.635, confirming that precise bounding box localization is more challenging than initial detection. The detailed IoU analysis (Table??) further shows a sharp decrease in AP after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {IoU}=0.75$$\end{document}.Primary Error Modes: The TIDE evaluation (Fig. 16) isolates the main sources of detection error. The most significant error type is Localization (Loc), representing the largest proportion of total error, which aligns with the observation from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document}. This implies that improving the regression branch to achieve higher positional accuracy should be a focus for future work. Missing Detections (Miss) is the second most common error, highlighting the difficulty in recalling all true instances, especially for small or highly clustered lesions. The model shows good performance in minimizing Classification (Cls) and Background (Bkg) false positives, suggesting the feature representations learned are generally discriminative (Fig. 17).Fig. 16TIDE Analysis of Error Types for the Proposed Framework (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Proposed-640}$$\end{document}). The pie chart and bar graphs illustrate the breakdown of major detection errors. Localization (Loc) errors (where the bounding box IoU is too low) constitute the largest failure mode, followed by Missing Detections (Miss). This analysis quantifies the model’s strengths in classification and its primary challenge in achieving highly precise localization. Scale Robustness (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}}$$\end{document}): The model demonstrates strong performance on medium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}} = 0.672$$\end{document}) and large (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}} = 0.720$$\end{document}) objects. However, performance significantly drops on small objects (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}} = 0.291$$\end{document}), suggesting that while the proposed MLFD and MDTA modules effectively enhance feature fusion, the detection of very subtle or distant lesions (small objects) remains the primary challenge. The corresponding recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {small}} = 0.329$$\end{document}) also confirms the difficulty in recovering these small instances. Scale Robustness (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}}$$\end{document}): The model demonstrates strong performance on medium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {medium}} = 0.672$$\end{document}) and large (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {large}} = 0.720$$\end{document}) objects. However, performance significantly drops on small objects (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}} = 0.291$$\end{document}), suggesting that while the proposed MLFD and MDTA modules effectively enhance feature fusion, the detection of very subtle or distant lesions (small objects) remains the primary challenge. The corresponding recall (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AR}_{\text {small}} = 0.329$$\end{document}) also confirms the difficulty in recovering these small instances. Localization Accuracy: The model achieves a competitive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{50}$$\end{document} of 0.837, indicating high accuracy at a loose Intersection over Union (IoU) threshold. However, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document} drops to 0.635, confirming that precise bounding box localization is more challenging than initial detection. The detailed IoU analysis (Table??) further shows a sharp decrease in AP after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {IoU}=0.75$$\end{document}. Localization Accuracy: The model achieves a competitive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{50}$$\end{document} of 0.837, indicating high accuracy at a loose Intersection over Union (IoU) threshold. However, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document} drops to 0.635, confirming that precise bounding box localization is more challenging than initial detection. The detailed IoU analysis (Table??) further shows a sharp decrease in AP after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {IoU}=0.75$$\end{document}. Primary Error Modes: The TIDE evaluation (Fig. 16) isolates the main sources of detection error. The most significant error type is Localization (Loc), representing the largest proportion of total error, which aligns with the observation from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document}. This implies that improving the regression branch to achieve higher positional accuracy should be a focus for future work. Missing Detections (Miss) is the second most common error, highlighting the difficulty in recalling all true instances, especially for small or highly clustered lesions. The model shows good performance in minimizing Classification (Cls) and Background (Bkg) false positives, suggesting the feature representations learned are generally discriminative (Fig. 17). Primary Error Modes: The TIDE evaluation (Fig. 16) isolates the main sources of detection error. The most significant error type is Localization (Loc), representing the largest proportion of total error, which aligns with the observation from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{75}$$\end{document}. This implies that improving the regression branch to achieve higher positional accuracy should be a focus for future work. Missing Detections (Miss) is the second most common error, highlighting the difficulty in recalling all true instances, especially for small or highly clustered lesions. The model shows good performance in minimizing Classification (Cls) and Background (Bkg) false positives, suggesting the feature representations learned are generally discriminative (Fig. 17). TIDE Analysis of Error Types for the Proposed Framework (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Proposed-640}$$\end{document}). The pie chart and bar graphs illustrate the breakdown of major detection errors. Localization (Loc) errors (where the bounding box IoU is too low) constitute the largest failure mode, followed by Missing Detections (Miss). This analysis quantifies the model’s strengths in classification and its primary challenge in achieving highly precise localization. Cross-species generalization evaluation To stringently validate the cross-species transferability and superior generalization capability of the proposed framework, an extensive transferability assessment was executed. Leveraging the foundational knowledge acquired from the primary rice leaf dataset, the model’s effectiveness was rigorously tested against two distinct and publicly accessible external datasets sourced from the Roboflow platform: Potato Leaf Diseases and Tomato Leaf Diseases. The Potato set, comprising \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {7,172}$$\end{document} images, features two major diseases (Early-Blight and Late-Blight). Conversely, the Tomato set, consisting of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {4,128}$$\end{document} images, covers eight different disease classifications (including Early Blight, Late Blight, Leaf Miner, Leaf Mold, Mosaic Virus, Septoria, Spider Mites, and Yellow Leaf Curl Virus). These auxiliary datasets were employed to evaluate the model’s robustness against varying leaf morphologies and visually disparate disease symptomatology, crucially without any subsequent fine-tuning or re-training. The resulting performance metrics from this cross-species study are documented in Fig. 18 and Table 8. As confirmed by the quantitative evidence, the proposed framework consistently achieved the peak \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} across both external validation sets, markedly surpassing leading lightweight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLO}$$\end{document} variants. Specifically, our model registered \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8412}$$\end{document} on the Tomato dataset and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.8867}$$\end{document} on the Potato dataset. This pronounced quantitative advantage confirms the architecture’s inherent capacity to effectively migrate learned features to novel plant pathology detection scenarios. Such heightened adaptability is directly attributed to the robust feature representation provided by our core architectural innovations (MLFD and MDTA). The successful validation of cross-species deployment firmly establishes this framework as a scalable and highly practical solution for unified disease monitoring within intelligent agricultural systems. Further visual analysis, including comparative attention heatmaps for this generalization experiment, is provided in Fig. 17.Fig. 17Comparison of heatmaps generated by different models on various datasets.Table 8Performance comparison of different models on various datasets (mAP@0.5:0.95).ModelRice leafTomato leafPotato leafYOLOv8n0.57410.81710.8754YOLOv10n0.56140.79840.8590YOLOv11n0.60850.82900.8755Ours0.62620.84120.8867The Tomato Leaf and Potato Leaf disease datasets were acquired from public domain resources. The dataset links are: Tomato Leaf Diseases: https://universe.roboflow.com/dyploma/tomato-leaf-diseases-4xa5i; Potato Leaf Diseases: https://universe.roboflow.com/final-year-project-9cf4g/potato-disease-balnv/dataset/2. Comparison of heatmaps generated by different models on various datasets. Table 8Performance comparison of different models on various datasets (mAP@0.5:0.95).ModelRice leafTomato leafPotato leafYOLOv8n0.57410.81710.8754YOLOv10n0.56140.79840.8590YOLOv11n0.60850.82900.8755Ours0.62620.84120.8867The Tomato Leaf and Potato Leaf disease datasets were acquired from public domain resources. The dataset links are: Tomato Leaf Diseases: https://universe.roboflow.com/dyploma/tomato-leaf-diseases-4xa5i; Potato Leaf Diseases: https://universe.roboflow.com/final-year-project-9cf4g/potato-disease-balnv/dataset/2. Performance comparison of different models on various datasets (mAP@0.5:0.95). The Tomato Leaf and Potato Leaf disease datasets were acquired from public domain resources. The dataset links are: Tomato Leaf Diseases: https://universe.roboflow.com/dyploma/tomato-leaf-diseases-4xa5i; Potato Leaf Diseases: https://universe.roboflow.com/final-year-project-9cf4g/potato-disease-balnv/dataset/2. Fig. 18Performance comparison of different models on various datasets (mAP@0.5:0.95). Performance comparison of different models on various datasets (mAP@0.5:0.95). Discussion This study successfully developed a novel methodology for rice leaf disease detection, building upon an enhanced iteration of the YOLOv11 architecture. The ultimate framework incorporates three principal architectural innovations: the Multi-Branch Large-Kernel Fusion Depthwise (MLFD) module, the Multi-Scale Dilated Transformer Attention (MDTA) module, and an efficient, lightweight Lo-Head detection head. These integrated components collectively overcome several critical challenges intrinsic to conventional detection techniques, including the complexities of field backgrounds, the difficulty of identifying minute lesions, and the need for computational efficiency in resource-constrained deployment environments. Analysis of core technical contributions and performance The successful integration of the proposed modules is key to the framework’s superior performance and balanced efficiency, as evidenced by the ablation study (Table 4).Feature Augmentation (MLFD and MDTA): The MLFD component utilizes its multi-branch topology and large convolutional kernels to enrich diverse feature representation and contextual information aggregation, a process critical for identifying disease characteristics under varying environmental conditions. Complementing this, the MDTA module further enhances the model’s capacity to adapt to complex, multi-scale disease manifestations by integrating attention mechanisms for effective fine-grained feature detection. This potent synergy is evidenced by the enhanced feature extraction capabilities observed in the module variants.Efficiency Optimization (Lo-Head): The Lo-Head component substantially reduces the computational burden of the model, achieving a highly optimized model size (5.6 MB) and maintaining a competitive inference speed (106.48 FPS). This lightweight design ensures the scheme is suitable for real-time applications where computational resources are limited.Overall Performance Gain: Experimental results show that the proposed framework network exhibits superior performance in terms of precision, recall, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document} compared to baseline models and other state-of-the-art detectors (Table 6). Especially in the critical \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} metric, the model improves from the baseline’s 0.6085 to 0.6262. The modest increase in model size (0.4 MB) for a significant performance gain validates the effectiveness and efficiency of the proposed component fusion. Feature Augmentation (MLFD and MDTA): The MLFD component utilizes its multi-branch topology and large convolutional kernels to enrich diverse feature representation and contextual information aggregation, a process critical for identifying disease characteristics under varying environmental conditions. Complementing this, the MDTA module further enhances the model’s capacity to adapt to complex, multi-scale disease manifestations by integrating attention mechanisms for effective fine-grained feature detection. This potent synergy is evidenced by the enhanced feature extraction capabilities observed in the module variants. Feature Augmentation (MLFD and MDTA): The MLFD component utilizes its multi-branch topology and large convolutional kernels to enrich diverse feature representation and contextual information aggregation, a process critical for identifying disease characteristics under varying environmental conditions. Complementing this, the MDTA module further enhances the model’s capacity to adapt to complex, multi-scale disease manifestations by integrating attention mechanisms for effective fine-grained feature detection. This potent synergy is evidenced by the enhanced feature extraction capabilities observed in the module variants. Efficiency Optimization (Lo-Head): The Lo-Head component substantially reduces the computational burden of the model, achieving a highly optimized model size (5.6 MB) and maintaining a competitive inference speed (106.48 FPS). This lightweight design ensures the scheme is suitable for real-time applications where computational resources are limited. Efficiency Optimization (Lo-Head): The Lo-Head component substantially reduces the computational burden of the model, achieving a highly optimized model size (5.6 MB) and maintaining a competitive inference speed (106.48 FPS). This lightweight design ensures the scheme is suitable for real-time applications where computational resources are limited. Overall Performance Gain: Experimental results show that the proposed framework network exhibits superior performance in terms of precision, recall, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document} compared to baseline models and other state-of-the-art detectors (Table 6). Especially in the critical \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} metric, the model improves from the baseline’s 0.6085 to 0.6262. The modest increase in model size (0.4 MB) for a significant performance gain validates the effectiveness and efficiency of the proposed component fusion. Overall Performance Gain: Experimental results show that the proposed framework network exhibits superior performance in terms of precision, recall, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}$$\end{document} compared to baseline models and other state-of-the-art detectors (Table 6). Especially in the critical \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} metric, the model improves from the baseline’s 0.6085 to 0.6262. The modest increase in model size (0.4 MB) for a significant performance gain validates the effectiveness and efficiency of the proposed component fusion. Implications for real-time precision agriculture The high speed (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {106.48 FPS}$$\end{document}) and lightweight nature of the proposed framework hold significant implications for real-time precision agriculture. Efficient disease detection is vital for improved disease prevention and timely crop management17. Our lightweight object detection approach provides the crucial low-latency necessary for deployment on mobile and edge devices, such as agricultural robots or drones, enabling instantaneous diagnosis in the field. Deployment Status Clarification: We have successfully deployed the proposed network onto an experimental pesticide robot setup within a controlled laboratory environment for preliminary functionality testing. However, we acknowledge that the network has not yet undergone comprehensive, long-term evaluation or deployment in a realistic, uncontrolled agricultural environment. This verification remains a critical step in future work to fully validate the scheme’s durability and practicality under real-world conditions. Furthermore, while our method focuses on object detection (bounding box localization), its success highlights the potential for combining high-speed detection with more fine-grained localization techniques. Specifically, low-latency semantic segmentation, which is crucial for determining the exact boundary and size of disease lesions, is highly complementary to fast detection42,43. Future integration with real-time semantic segmentation methods will allow agricultural robots not only to quickly locate diseased areas but also to precisely measure disease severity and target localized treatments, maximizing yield and minimizing pesticide use. Model interpretability and enhancing trustworthiness Model interpretability is a critical factor in deploying AI systems in real-world agricultural settings, where human trust and verification are essential for correct diagnoses. Our current work utilized Grad-CAM to generate heatmaps (Fig. 14), visually confirming that the proposed framework focuses on relevant lesion regions, thereby providing a foundational level of model explanation. However, the insights provided by gradient-based methods like Grad-CAM are localized and may not fully explain the model’s global decision-making process. To further enhance the trustworthiness and transparency of the proposed network, future work should explore the integration of model-agnostic explanation techniques, as suggested by recent research in agricultural XAI33–36:LIME (Local Interpretable Model-agnostic Explanations): LIME focuses on explaining the model’s predictions locally by approximating the complex model with a simpler, interpretable one around the specific instance. This would be valuable for understanding the influence of individual pixels or small patches on a single disease prediction for a specific rice leaf.SHAP (SHapley Additive exPlanations): SHAP provides a unified framework for interpreting predictions by assigning each feature (or pixel) an importance value (Shapley value) for a particular prediction. Utilizing SHAP would allow us to quantify the contribution of specific visual features (e.g., color, texture, shape) to the final disease detection score across the entire dataset.The potential integration of LIME or SHAP would allow for more robust verification of model decision boundaries and provide stronger evidence that the network is learning genuine pathological patterns. LIME (Local Interpretable Model-agnostic Explanations): LIME focuses on explaining the model’s predictions locally by approximating the complex model with a simpler, interpretable one around the specific instance. This would be valuable for understanding the influence of individual pixels or small patches on a single disease prediction for a specific rice leaf. LIME (Local Interpretable Model-agnostic Explanations): LIME focuses on explaining the model’s predictions locally by approximating the complex model with a simpler, interpretable one around the specific instance. This would be valuable for understanding the influence of individual pixels or small patches on a single disease prediction for a specific rice leaf. SHAP (SHapley Additive exPlanations): SHAP provides a unified framework for interpreting predictions by assigning each feature (or pixel) an importance value (Shapley value) for a particular prediction. Utilizing SHAP would allow us to quantify the contribution of specific visual features (e.g., color, texture, shape) to the final disease detection score across the entire dataset. SHAP (SHapley Additive exPlanations): SHAP provides a unified framework for interpreting predictions by assigning each feature (or pixel) an importance value (Shapley value) for a particular prediction. Utilizing SHAP would allow us to quantify the contribution of specific visual features (e.g., color, texture, shape) to the final disease detection score across the entire dataset. Limitations and future work Despite the commendable performance achieved, several avenues for further research and practical enhancement remain regarding the model’s deployment and application scope:Enhanced Environmental Resilience: Although the model yielded strong outcomes in our controlled experimental setup, its deployment to diverse, real-world agricultural environments–characterized by unmanaged variables like fluctuating light, leaf overlap (shading), complex soil backgrounds, and dynamic camera angles–presents ongoing challenges. Subsequent investigations will prioritize specialized data augmentation or domain adaptation methodologies to bolster the model’s robustness against these environmental variabilities.Cross-Crop Versatility and Dataset Diversity: While we successfully demonstrated notable generalization capability across a limited selection of other plant species, the core training of the model was centered on rice leaf pathology. Expanding its utility to other vital agricultural commodities (e.g., coffee, chili peppers, or maize) will necessitate support from broader datasets and will require additional optimization of the underlying detection module to effectively manage novel lesion morphologies.Hardware Deployment Optimization: When transitioning this network for use on resource-constrained edge devices (such as autonomous agricultural robots or drones), further hardware-aware optimization beyond the initial Lo-Head module will become essential. This scope includes the exploration of techniques like network quantization and pruning to minimize inference latency while simultaneously preserving the highest feasible detection accuracy for real-world field operations. Enhanced Environmental Resilience: Although the model yielded strong outcomes in our controlled experimental setup, its deployment to diverse, real-world agricultural environments–characterized by unmanaged variables like fluctuating light, leaf overlap (shading), complex soil backgrounds, and dynamic camera angles–presents ongoing challenges. Subsequent investigations will prioritize specialized data augmentation or domain adaptation methodologies to bolster the model’s robustness against these environmental variabilities. Enhanced Environmental Resilience: Although the model yielded strong outcomes in our controlled experimental setup, its deployment to diverse, real-world agricultural environments–characterized by unmanaged variables like fluctuating light, leaf overlap (shading), complex soil backgrounds, and dynamic camera angles–presents ongoing challenges. Subsequent investigations will prioritize specialized data augmentation or domain adaptation methodologies to bolster the model’s robustness against these environmental variabilities. Cross-Crop Versatility and Dataset Diversity: While we successfully demonstrated notable generalization capability across a limited selection of other plant species, the core training of the model was centered on rice leaf pathology. Expanding its utility to other vital agricultural commodities (e.g., coffee, chili peppers, or maize) will necessitate support from broader datasets and will require additional optimization of the underlying detection module to effectively manage novel lesion morphologies. Cross-Crop Versatility and Dataset Diversity: While we successfully demonstrated notable generalization capability across a limited selection of other plant species, the core training of the model was centered on rice leaf pathology. Expanding its utility to other vital agricultural commodities (e.g., coffee, chili peppers, or maize) will necessitate support from broader datasets and will require additional optimization of the underlying detection module to effectively manage novel lesion morphologies. Hardware Deployment Optimization: When transitioning this network for use on resource-constrained edge devices (such as autonomous agricultural robots or drones), further hardware-aware optimization beyond the initial Lo-Head module will become essential. This scope includes the exploration of techniques like network quantization and pruning to minimize inference latency while simultaneously preserving the highest feasible detection accuracy for real-world field operations. Hardware Deployment Optimization: When transitioning this network for use on resource-constrained edge devices (such as autonomous agricultural robots or drones), further hardware-aware optimization beyond the initial Lo-Head module will become essential. This scope includes the exploration of techniques like network quantization and pruning to minimize inference latency while simultaneously preserving the highest feasible detection accuracy for real-world field operations. Conclusion In this work, we successfully introduced a novel and effective detection framework for the prompt identification of rice leaf diseases in natural field environments, based on an enhanced lightweight architecture. The framework incorporates three principal innovations: the Multi-Branch Large-Kernel Fusion Depthwise (MLFD) module for robust feature enhancement, the Multi-Scale Dilated Transformer Attention (MDTA) module for improved environmental resilience, and the lightweight Lo-Head detection head for computational efficiency. This specialized design outperformed all compared architectures, recording a peak \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {0.6262}$$\end{document}–a gain of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {1.77}$$\end{document} percentage points over the baseline (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {YOLOv11n}$$\end{document}). Crucially, the architecture retains high efficiency, achieving \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {106.48 FPS}$$\end{document} with a compact model size of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {5.6 MB}$$\end{document}, validating its fitness for deployment on resource-constrained edge devices. Moreover, the framework exhibited significant cross-species transferability, attaining a high \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {mAP}@0.5:0.95$$\end{document} on potato (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {88.67}$$\end{document}) and tomato (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {84.12}$$\end{document}) diseases, highlighting its potential as a unified and scalable solution for multi-crop monitoring in smart farming. Despite the demonstrated superiority, TIDE analysis pinpoints that bolstering localization accuracy and detection performance for small targets (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_{\text {small}} = 0.291$$\end{document}) constitute the foremost challenges. Future investigations will concentrate on extending the model’s environmental robustness across a wider array of crop species (including coffee, pepper, and corn) and optimizing the localization branch through detailed error analysis. Ultimately, our focus will shift toward optimizing the framework for practical field deployment and validation onto agricultural robots, thereby facilitating real-time detection and early intervention in uncontrolled field settings to maximize the network’s utility in intelligent agricultural systems.