DAN-YOLO-L: a lightweight YOLO approach for efficient infrared small object detection from UAV perspectives Authors: Kai Feng, Guojun Lin, Tong Lin, Zhongqiang Luo, Run Ma Published: 2026 ABSTRACT AbstractSmall object detection in drone aerial photography faces challenges such as small scales, inconspicuous features, and complex backgrounds. To address these issues, this paper proposes an improved detection model based on the YOLO framework. First, a multi-channel feature extraction module—Deconvolutional Network combined with the C3 module (deconv-c3k2)—is designed to enhance feature extraction and multi-scale representation capabilities. Second, an enhanced Auxiliary Head detection module is introduced to improve feature interaction and collaboration across different levels. Concurrently, the NWD-Inner-CIoU loss function is adopted to mitigate the impact of IoU on small target localization offset, thereby boosting detection accuracy. To meet real-time embedded deployment requirements, an L1 pruning strategy is employed to reduce the model parameter size. Experimental results demonstrate that the proposed method significantly outperforms baseline models on the HIT-UAV dataset, achieving 81.9% on mAP@0.5 and 51.4% on mAP@0.5:0.95, respectively. Inference speed increases by 8.32%, while parameter count decreases by 20.8%. Computational load was reduced by 0.36 GFLOPs, and model size shrinks to 79.2% of the original. Furthermore, experiments on the IRay Infrared Dataset validate the method’s generalization capability. Overall, the proposed approach demonstrates distinct advantages in detection accuracy, real-time performance, and lightweight design, while exhibiting stability and practical value.Keywords: Small object detection, UAV infrared images, DAN-YOLO-L, NWD-Inner-CIoU lossSubject terms: Engineering, Mathematics and computing, Optics and photonics Introduction With the rapid advancement of unmanned aerial vehicle (UAV) technology, UAVs are increasingly applied in fields such as remote sensing, traffic monitoring, and disaster relief1–3. Particularly in complex environments, detecting small targets within UAV aerial imagery poses significant challenges.According to standard benchmarks such as VisDrone and MS COCO, small targets typically exhibit minimal pixel dimensions (e.g., less than 16× 16 or 32× 32 pixels)4,5, resulting in blurred features and insufficient information. Consequently, existing detection methods often demonstrate low accuracy and high false positive rates when handling such targets. To cope with this drawback, enhancing the accuracy of small target detection while reducing computational resource consumption has emerged as a key research problem. In recent years, with the widespread application of deep learning technologies, object detection methods based on convolutional neural networks (CNNs)6 have made significant progress. The YOLO series of models7, in particular, has demonstrated balance between accuracy and computational complexity. Despite its superior balance between detection speed and accuracy in general scenarios, the YOLO framework, like most single-stage detectors, still encounters specific performance bottlenecks when handling small objects in complex environments, particularly regarding feature preservation and localization stability. First, traditional feature extraction methods often neglect multi-scale feature fusion, leading to suboptimal model performance across different scales of small objects. Second, positioning errors in small objects cause instability in Intersection over Union (IoU)8 calculations. Consequently, traditional IoU-based loss functions frequently fail to meet the demands of small object detection, resulting in increased training difficulty and reduced convergence speed. To address these challenges in small object detection, this paper proposes an innovative detection framework with significant improvements in four key aspects: Multi-scale feature fusion enhancement: To overcome feature loss caused by scale differences in traditional detection frameworks, this paper integrates deconvolution9 with the C3k2 architecture to enhance multi-scale feature10 extraction and fusion capabilities. This integration not only improves feature representation of small objects across scales but also effectively enhances the model’s sensitivity to low-resolution small targets. This method is particularly suitable for detecting small objects in drone aerial imagery, significantly improving detection accuracy and robustness. Multi-scale feature fusion enhancement: To overcome feature loss caused by scale differences in traditional detection frameworks, this paper integrates deconvolution9 with the C3k2 architecture to enhance multi-scale feature10 extraction and fusion capabilities. This integration not only improves feature representation of small objects across scales but also effectively enhances the model’s sensitivity to low-resolution small targets. This method is particularly suitable for detecting small objects in drone aerial imagery, significantly improving detection accuracy and robustness. Improved auxiliary head: Addressing the limitations of traditional detection heads in small object detection, this paper proposes an improved Auxiliary Head11. By optimizing multi-level feature fusion strategies, this detection head more efficiently leverages low-level detail information and high-level semantic information. It notably enhances model accuracy, especially for small objects and complex backgrounds. Compared to the standard YOLO detection head, the improved Auxiliary Head boosts the network’s performance in multi-scale small object detection, enhancing both object localization and classification capabilities. Improved auxiliary head: Addressing the limitations of traditional detection heads in small object detection, this paper proposes an improved Auxiliary Head11. By optimizing multi-level feature fusion strategies, this detection head more efficiently leverages low-level detail information and high-level semantic information. It notably enhances model accuracy, especially for small objects and complex backgrounds. Compared to the standard YOLO detection head, the improved Auxiliary Head boosts the network’s performance in multi-scale small object detection, enhancing both object localization and classification capabilities. NWD-Inner-CIoU loss function: To address the excessive sensitivity to positional bias in traditional IoU loss functions for small object detection, this paper introduces the NWD-Inner-CIoU12 loss function. By combining Normalized Wasserstein Distance (NWD) with an Inner Auxiliary Box mechanism, this approach significantly improves localization accuracy for small objects. Compared to conventional IoU, NWD-Inner-CIoU better measures the distributional similarity between bounding boxes, demonstrating superior robustness in small-object and low-resolution scenarios while reducing false negatives and false positives. NWD-Inner-CIoU loss function: To address the excessive sensitivity to positional bias in traditional IoU loss functions for small object detection, this paper introduces the NWD-Inner-CIoU12 loss function. By combining Normalized Wasserstein Distance (NWD) with an Inner Auxiliary Box mechanism, this approach significantly improves localization accuracy for small objects. Compared to conventional IoU, NWD-Inner-CIoU better measures the distributional similarity between bounding boxes, demonstrating superior robustness in small-object and low-resolution scenarios while reducing false negatives and false positives. L1 pruning strategy: To address the dual challenges of real-time performance and computational resource constraints on UAV platforms, this paper introduces the L1 pruning strategy13. By pruning unimportant connections within the model, L1 significantly reduces the number of parameters and computational load while preserving detection accuracy. Unlike conventional pruning methods, L1 optimizes inference speed without compromising model performance, enabling it to meet the stringent computational and temporal demands of embedded systems. L1 pruning strategy: To address the dual challenges of real-time performance and computational resource constraints on UAV platforms, this paper introduces the L1 pruning strategy13. By pruning unimportant connections within the model, L1 significantly reduces the number of parameters and computational load while preserving detection accuracy. Unlike conventional pruning methods, L1 optimizes inference speed without compromising model performance, enabling it to meet the stringent computational and temporal demands of embedded systems. The remainder of this paper is organized as follows: Section “Related work” reviews prior work in small object detection and identifies limitations of existing approaches. Section “Proposed method” details the innovative design of the proposed enhanced detection framework. Section “Experimental results” provides the experimental setup and methodology, comparing our approach with several existing baseline models and the latest state-of-the-art (SOTA) methods. Experimental results demonstrate the superiority of our proposed method in terms of accuracy, real-time performance, and model lightweighting. The robustness of the method is explored, showcasing stable performance in complex scenarios. Finally, “Conclusions” section concludes the research work and outlines future research directions for small object detection in the field of remote sensing. Related work Object detection, as a significant research topic in computer vision, has undergone a developmental journey from shallow to deep approaches. Early methods primarily relied on manually designed features and traditional classifiers, such as Haar features combined with Adaboost, HOG with SVM14–16, and sliding window detection. While these methods advanced object detection applications at the time, their limited feature representation capabilities made them ill-suited for complex backgrounds and small objects. Additionally, their computational efficiency was relatively low, failing to meet real-time requirements. With the rise of deep learning, detection frameworks based on convolutional neural networks gradually replaced traditional approaches. Two-stage models like R-CNN17, Fast R-CNN18, and Mask R-CNN19 achieved significant accuracy improvements by performing detection in two steps: candidate region generation and classification/regression. However, these methods generally incurred high computational overhead and slow inference speeds, making them unsuitable for drone aerial photography tasks requiring rapid response. To address this speed bottleneck, researchers proposed one-stage detectors like SSD20 and the YOLO series. YOLO transforms detection into a regression problem, directly predicting categories and bounding boxes on feature maps, greatly enhancing real-time performance. Through iterations from the original YOLO to the latest YOLOv11, the network has continuously optimized its backbone architecture, feature fusion, and detection head design, achieving an optimal balance between speed and accuracy. Consequently, it has gradually become a key benchmark framework for small object detection in UAV applications. Researchers have also introduced targeted improvements for small object detection. For instance, Dual-YOLO21 enhances multi-scale representation through dual-path and feature fusion mechanisms. SOD-YOLO22 introduced cross-domain attention modules to enhance small object discrimination in aerial imagery. IPD-YOLO23 designed novel detection heads and loss functions for infrared scenarios, improving localization accuracy for weakly featured targets. NOC-YOLO24 attempted to balance lightweight architecture with detection precision, demonstrating feasibility in infrared small-object vehicle detection.Recently, Transformer-based approaches have also demonstrated a growing impact on small object detection tasks. For instance, recent studies, such as the contrast-aware and uncertainty-optimized transformer model, have shown that Transformer architectures can significantly improve small object representation and localization in complex high-resolution scenes25. Similarly, recent advancements have explored lightweight transformer models incorporating adaptive rotational convolutions to specifically address aerial object detection, further demonstrating the potential of self-attention mechanisms in extracting robust features from complex UAV perspectives26. However, while Vision Transformers excel in capturing global context and fine-grained representations, their high computational complexity and substantial memory footprint often limit their practical deployment on resource-constrained UAV platforms that require real-time inference. Therefore, lightweight CNN architectures, particularly the YOLO series, remain a more feasible and highly efficient framework for UAV-based small object detection. These efforts lay the groundwork for small object detection, yet challenges persist in complex environments–including insufficient accuracy, unstable features, and the difficulty of balancing lightweight design with performance. Infrared imaging, while effective for nighttime and adverse weather conditions, often suffers from low signal-to-noise ratios and a lack of texture information. As highlighted in recent studies on infrared shape and motion modeling27,28, small targets in these images are frequently submerged in complex, cluttered backgrounds, such as clouds, buildings, or sea waves. These factors make it extremely difficult for conventional networks to distinguish genuine targets from high-intensity noise or thermal artifacts Proposed method This study builds upon the YOLOv11n framework to address the challenges of insufficient feature representation and high model complexity in small object detection. By refining the network architecture and optimization strategies, we developed a lightweight and efficient detection model, DAN-YOLO-L, whose structural diagram is shown in Fig. 1. During feature extraction, the C3k2 module is combined with a deconvolution (Deconv) structure to enhance feature recovery and detail representation, mitigating feature loss in small objects caused by downsampling. The detection component employs an auxiliary detection head (Aux Head), introducing additional supervisory signals at intermediate layers to promote feature optimization and improve detection stability. For the loss function, NWD-Inner-CIoU is introduced. By integrating normalized Wasserstein distance with the CIoU metric, it simultaneously accounts for differences in object center position, scale, and shape, thereby improving localization accuracy for small targets. Finally, L1 sparse pruning removes redundant channels, effectively reducing model parameters and computational overhead to achieve lightweight deployment. Overall framework of DAN-YOLO-L. Feature extraction module improvement: C3k2_DEConv To enhance the network’s ability to represent multi-scale objects and fine-grained boundaries, this study introduces the C3k2_DEConv module during the feature extraction stage of the backbone network. Building upon the C3 architecture from the YOLO series, this module integrates deep deconvolution with a two-stage kernel expansion mechanism (k2),as shown in Fig. 2. This approach achieves high-resolution semantic restoration and structural detail enhancement while maintaining computational efficiency. Drawing inspiration from DeconvNet29, This method progressively reconstructs object morphology through deconvolution and unpooling, effectively addressing the semantic loss that occurs during traditional CNN upsampling. We designed a novel module–C3k2_DEConv. Figure 3 illustrates the overall network topology to visually demonstrate the proposed C3k2_DEConv module’s structure and information flow. It can be observed that the backbone branch handles deep semantic feature extraction, while the deconvolution branch restores spatial details through layer-by-layer upsampling and structural reconstruction, thereby establishing an effective balance between semantic abstraction and spatial resolution. Specifically, the C3k2_DEConv module embeds a deconvolution branch within the bottleneck layer of the traditional C3 architecture to achieve layer-wise refinement from the semantic space to the spatial domain. The overall network structure can be viewed as: Here, represents feature concatenation and fusion operation. This structure ensures that high-level abstraction of semantic expression and local precision of spatial representation are synergistically optimized at the same stage. In the backpropagation, the feature map is re-aligned in space by performing convolution operations. Through the convolution operation, it is refined by the feature map and the convolutional kernel W, where the convolution operation is represented as: The symbol “*” represents the convolution operation, and represents the transposed convolution kernel, which is used to map low-dimensional features into higher-dimensional feature space. To ensure spatial consistency, the operation is applied to output feature maps, ensuring that the size aligns with the boundaries. Meanwhile, the depooling layer uses the index matrix S from the max pooling stage to restore the activation values to their original positions. This mechanism effectively restores the geometric consistency of spatial structures, enabling high-level semantic information to retain precise localization during the upsampling process. During the fusion stage, the C3k2_DEConv module, employs an adaptive channel weighting strategy30 to achieve a dynamic balance between semantic and structural information. The overall output can be expressed as: In this study, the weighting factor α was empirically set to 0.5 to provide an equal balance between global semantic expressions and local structural details. This weighting strategy not only enhances the response intensity in small target regions but also effectively suppresses background noise.The DEConv module comprises four differential convolutions (DC) and one standard convolution (VC), with its network architecture illustrated in Fig. 4. First, the input image undergoes preliminary feature extraction via the standard convolution (VC). Subsequently, four differential convolutions (CDC, ADC, VDC, and HDC) further process the image to capture detailed features across different orientations and scales. These difference convolution operators are formulated based on the principles of explicitly modeling gradient-level feature variations31. Specifically, rather than performing standard local feature aggregation, these modules compute explicit differences: CDC (Channel Differential Convolution) captures inter-channel feature discrepancies to enhance semantic discrimination; VDC (Vertical Differential Convolution) and HDC (Horizontal Differential Convolution) aggregate local spatial intensity differences along the y-axis and x-axis respectively, effectively acting as learnable edge detectors. Furthermore, ADC (Axial Differential Convolution) integrates these directional gradients to capture more complex boundary topologies. These operations work synergistically. These operations work synergistically to significantly enhance DEConv’s capability in handling image details, particularly in complex small object detection scenarios, where they markedly improve feature extraction accuracy. In summary, the C3k2_DEConv module follows a three-step process: (1) compression of C3 backbone features, (2) deconvolution-based detail reconstruction, and (3) adaptive fusion of the reconstructed details. This module significantly enhances feature structural integrity and multi-scale robustness while maintaining inference efficiency. It effectively improves boundary prediction accuracy in complex scenes and provides more refined, semantically consistent high-resolution features for subsequent detection and segmentation tasks. C3k network architecture. The parameter n denotes the number of Bottleneck_DEConv layers, set to 2 in this paper. C3k2_DEConv network architecture. DEConv architecture. To further enhance the model’s multi-scale detection performance, this study designs an improved detection head that integrates a multi-scale semantic enhancement mechanism based on the high-level semantic features of the backbone network. The overall architecture comprises a primary detection head and an auxiliary detection head (Aux Head). The primary detection head handles global object detection, while the auxiliary detection head locally enhances shallow, fine-grained features. This dual approach achieves semantic consistency enhancement and gradient coordination. The overall structural difference from conventional detection models is illustrated in Fig. 5, which compares the traditional single-detector configuration with the improved architecture introduced in this paper featuring an auxiliary detector. Conventional detection models perform detection solely on high-level semantic features, lacking explicit supervision over mid-to-shallow-level features. In contrast, the proposed DetectAux architecture introduces auxiliary detection branches at the same multi-scale feature layers (8×, 16×, 32×) as the main detection head. This provides additional supervision signals and gradient feedback exclusively during training, thereby enhancing feature representation capabilities and mitigating the vanishing gradient problem. During the feature extraction stage of the backbone network, five feature maps at different levels are generated, corresponding to network layer indices of 10, 13, 16, 19, and 22 respectively. These feature layers contain multi-scale information ranging from shallow textures to deep semantics. To fully leverage their complementary properties, this paper adopts a “six-layer feature map–three detection heads” architecture. After integrating the multi-layer features through a fusion module, they are fed into three detection heads, corresponding to small, medium, and large object detection tasks. The fusion relationship can be expressed as: Here, denotes the feature map extracted from the i-th layer of the network (e.g., represents the feature map from layer 16). The variables , , and represent the fusion results of three different scales (small, medium, and large). The fusion function incorporates upsampling, channel compression, and feature concatenation operations to achieve spatial and semantic alignment between features of different resolutions. This structure mitigates the fragmentation of single-layer input information while enabling the collaborative expression of shallow spatial features and deep semantic features. To achieve gradient-cooperative optimization between the primary and auxiliary detector heads, this paper designs the DetectAux network architecture shown in Fig. 6. This structure employs an Assigner module to match ground truth (GT) with prediction results while simultaneously computing losses for both detector heads, enabling bidirectional information flow constraints. During training, the auxiliary detector head provides additional supervisory signals to stabilize gradient propagation and enhance the discriminative capability of shallow features, while the main detector head handles the final prediction task. The primary detector head adopts the single-stage detector design philosophy, comprising three prediction branches corresponding to different scales: small, medium, and large object detection tasks. Each branch consists of a convolutional stacking module, a Batch Normalization layer, and a nonlinear activation function, outputting class probabilities, object confidence scores, and bounding box regression offsets. Regarding feature input, the main detector head not only receives multi-scale features from the fusion module but also acquires supplementary information from adjacent layers through cross-layer connections. This enhances cross-scale semantic consistency and feature representation completeness. Its prediction output can be formally expressed as: Among them, and are learnable parameters, and represents a nonlinear activation function. This architecture ensures the model can simultaneously focus on local details and global context, thereby enhancing small object detection performance. To mitigate the issue of gradient vanishing in the main detection head and enhance the recognition capability of shallow-layer features, this paper introduces an Auxiliary Detection Head (Aux Head) during the training phase. This module is activated only during training, constructing auxiliary gradient paths through additional supervisory signals. This enables multi-source gradient feedback and collaborative optimization across feature levels. Among them, , , and represent convolution kernels of different sizes, and denotes the Sigmoid activation function. Its output is defined as: Primary and auxiliary detection heads employ joint loss optimization: where represents the loss from the primary (lead) detection head, represents the loss from the auxiliary detection head, and α is a weighting coefficient used to balance the contribution of the auxiliary supervision signal. This closed-loop design, named “six-layer input, three-head output, and auxiliary supervision”, significantly improves gradient stability and small-object detection performance while maintaining high inference efficiency. Structural comparison between standard multi-scale detection and the proposed dual-head framework. DetectAux network architecture diagram. Loss function optimization: NWD-Inner-CIoU loss function for small object detection To further enhance detection accuracy and regression stability, this study introduces a combined NWD-Inner-CIoU optimization scheme within the existing detection framework. Traditional IoU-based losses (e.g., GIoU, DIoU, CIoU) are overly sensitive to positional offsets in small objects and yield zero gradients when predicted boxes do not intersect or completely enclose ground-truth boxes.To address these issues, the Normalized Gaussian Wasserstein Distance (NWD) is adopted as a scale-normalized similarity metric, providing smooth and consistent gradients under varying object sizes and displacements.In parallel, the Inner-CIoU loss refines the geometric alignment between anchor boxes and target boxes by introducing an inner-box constraint that focuses optimization on the most representative overlap regions.This dual-loss design allows NWD to improve metric stability while Inner-CIoU enhances localization precision, particularly for small or densely packed targets. Figure 7 illustrates the geometric mechanism of the proposed Inner-CIoU loss, showing how inner and outer bounding regions are defined for both targets and anchors. The purple boxes correspond to the target-related regions (ground-truth and its corresponding anchor box), while the orange boxes denote their inner counterparts used to constrain relative positional alignment. During training, the proposed Inner-CIoU mechanism refines the traditional CIoU formulation by introducing an inner bounding region within both the target and anchor boxes.This inner region focuses the regression process on the core overlap area, allowing effective gradient propagation even when boxes are partially misaligned or non-overlapping. Guided by this geometry-aware constraint, the adaptive anchor–target assignment achieves more stable and precise localization.When combined with the NWD term, the overall framework maintains smooth differentiability and enhanced robustness across various object scales. Inner-CIoU loss function. According to the geometric relationship shown in Fig. 7, the coordinates of the inner auxiliary bounding box are defined by scaling the original bounding box while maintaining the same center point: where denotes the center coordinates, w and h represent the width and height of the bounding box, and ratio is the scaling factor illustrated by the orange boxes in Fig. 7. Subsequently, the Inner-CIoU loss is formulated to focus the regression on this core overlap area: First, model the predicted box and the ground-truth box as two-dimensional Gaussian distributions: where: , can be defined similarly as the mean and the covariance of the ground-truth box. This modeling approach reflects the importance weight distribution of bounding box centers, enabling the transformation of box similarity into a problem of calculating distances between distributions. The second-order Wasserstein distance between two distributions can be expressed as: and further normalized into a similarity form: The constant C is an empirically determined scale factor closely related to the dataset’s average object dimensions, ensuring that the distance metric smoothly maps to the (0, 1] interval. In our implementations for both the HIT-UAV and IRay datasets, following standard practices in small object detection, C was empirically set to 12.8 to guarantee optimal gradient propagation and robust scale normalization. The NWD loss is defined as: To leverage the scale-robustness of NWD and the precise geometric constraints of Inner-CIoU, we propose the combined NWD-Inner-CIoU regression loss. The overall bounding box regression loss () is defined as the weighted sum of both terms: Finally, this combined regression loss is integrated with the classification loss () and the distribution focal loss () to form the overall optimization objective for the network: Through this joint optimization formulation, the model maintains stable gradients even in regions without intersections, effectively avoiding the degradation issues of traditional IoU-based losses in small object detection. Combined with the C3k2_DEConv module and multi-scale detection heads, the model exhibits improved smoothness and convergence during gradient propagation, achieving higher localization accuracy and detection robustness, particularly in complex scenes. L1 pruning strategy To reduce model complexity and inference latency while maintaining detection accuracy, this paper introduces a channel pruning strategy based on L1-norm regularization during the late stages of model training. This approach structurally sparsifies the C3k2_DEConv module and the detection head’s convolutional layers. Its core principle involves automatically identifying and removing channels or convolutional kernels with minimal contribution during weight optimization, thereby achieving parameter compression and computational acceleration. By introducing an L1 regularization term on the weights in the convolutional layer, the weight distribution tends toward sparsity: where is the sparsity coefficient, used to balance detection accuracy and compression ratio. The L1 norm term tends to push some channel weights toward zero, thereby identifying “ineffective channels” without significantly compromising the representation of key features. During the pruning phase, channels are sorted by their absolute weight values in ascending order and a threshold τ is set. When , channel i is deemed redundant and removed. The pruning operation can be formalized as: Through short-term fine-tuning, we recover performance losses caused by pruning. The final model exhibits significantly reduced parameter count and FLOPs, while achieving an average inference speed improvement of approximately 25 Compared to traditional unstructured pruning, this method achieves structured reduction along the channel dimension, enabling direct compatibility with mainstream inference frameworks without requiring sparse matrix operations. Furthermore, by integrating the previously designed multi-scale detection heads with the NWD loss function, the L1-pruned model retains strong feature representation capabilities and detection stability, achieving a balance between compact structure and robust performance. Experimental dataset description To validate the detection performance of the proposed model under diverse scenarios, this paper selected two representative public datasets: HIT-UAV32 and the IRAY Infrared Dataset.Both datasets focus on infrared small-object detection but offer different perspectives: the former emphasizes high-altitude UAV aerial thermal imagery, while the latter focuses on multi-scale target detection in diverse ground-based infrared surveillance scenarios. Together, they provide a solid basis for evaluating the model’s generalization performance under complex thermal conditions, jointly providing a solid basis for evaluating the model’s generalization performance under complex conditions. (1) The HIT-UAV dataset, proposed by Harbin Institute of Technology, aims to advance research in infrared small object detection. It is constructed from infrared video sequences captured by a low-altitude drone platform, covering diverse complex scenarios including campuses, roads, grasslands, built-up areas, and nighttime environments. A total of 2898 thermal infrared image frames with a resolution of 384 × 288 were extracted from the raw videos, with precise bounding box annotations provided for each target. Target types primarily include pedestrians, vehicles, and low-altitude aircraft, with sizes typically occupying less than 1% of the image area, demonstrating the characteristic features of small objects. As illustrated in Fig. 8a, the HIT-UAV dataset’s high background noise and low signal-to-noise ratio characteristics impose stringent demands on the multi-scale modeling capabilities of detection algorithms. Consequently, it is widely adopted for evaluating infrared small target detection performance. (2) The IRay Infrared Dataset, released by China’s IRay Optoelectronics, serves as a universal dataset for validating infrared intelligent perception algorithms. It encompasses surveillance images across multiple scenarios under both daytime and nighttime conditions, covering typical thermal target categories such as pedestrians, vehicles, and small aerial targets. The IRay dataset analysis diagram is shown in Fig. 8b. The dataset comprises 11,045 infrared images, with 8753 used for training and 2292 for testing. All images share a uniform resolution of 640 × 512 pixels. Significant challenges exist within the dataset, including thermal radiation interference, edge blurring, and target occlusion, making it a representative benchmark for validating model robustness and infrared feature recognition capabilities. During preprocessing, images undergo grayscale normalization and histogram-based heat enhancement to mitigate sensor noise effects. To ensure experimental fairness and reproducibility, all experiments follow identical training strategies with input images uniformly resized to 640 × 640. Training and validation data were split in an 8:2 ratio, with the objective function optimized using a joint loss formulation. Data analysis chart. Experimental results Training environment setup The experimental environment is based on the Windows 11 operating system, with hardware configurations including an Intel Core i5-14600KF processor, 32 GB of RAM, and an NVIDIA GeForce RTX 4060 Ti graphics card with 16 GB of VRAM. The experiment employed Python 3.10.0 and the PyTorch 2.2.2 (cu121) framework, with CUDA version 11.3. Relevant parameter settings are detailed in Table 1.While the training batch size was set to 16, all inference performance metrics, including FPS, were measured with a batch size of 1 and FP32 precision to simulate real-time UAV deployment. Table 1Experimental parameter settings.SettingsParametersSettingsParametersImages640 × 640Patience20Batch16Momentum0.937Epochs200lr00.01Workers8lrf0.01 Experimental parameter settings. Evaluation indicators To comprehensively evaluate the performance of the improved model in small object detection tasks, this paper constructs an evaluation framework across three dimensions: detection accuracy, classification capability, and computational efficiency. Multiple metrics including mAP, precision, recall, parameter count, and GFLOPs are employed for quantitative analysis. (1) Mean Average Precision (mAP) In the field of object detection, the average precision (AP) for a single class is typically measured by the area under the precision-recall curve for that class. It is defined as the area under the precision-recall curve for target class c. Its mathematical expression is: Calculate the arithmetic mean of the AP values across all categories to obtain the overall performance metric mAP: In the formula, C denotes the total number of categories. To evaluate detector performance across different levels of localization stringency, this paper reports two commonly used metrics: mAP@0.5 (matches determined at IoU = 0.5) reflects basic matching capability; mAP@0.5:0.95 (calculated by averaging 10 thresholds taken at 0.05 intervals within to more rigorously evaluate localization precision. (2) Precision and Recall Precision and Recall are used to evaluate the accuracy and coverage of detection results, respectively, with the following calculation formulas: Here, TP, FP, and FN represent the counts of true positives, false positives, and false negatives, respectively. A high Precision indicates fewer false positives, while a high Recall indicates fewer missed detections. The relative changes between these metrics reveal the trade-offs in the model’s performance across different thresholds or scenarios. (3) Model Parameters and Computational Complexity Considering practical deployment scenarios (especially edge or real-time systems), we simultaneously evaluate both the storage and computational overhead of models: Parameter size (Params) reflects the storage required for model parameters (typically measured in millions, M), while GFLOPs describe the approximate floating-point operations performed during a single forward inference pass, defined as follows: The values of parameters and GFLOPs provide an intuitive indication of a model’s adaptability in resource-constrained environments–lower values are more conducive to real-time inference and resource conservation. In summary, this study employs a multi-metric approach to evaluate models across multiple dimensions, from detection performance to resource efficiency, aiming to provide a more comprehensive reference for model selection and optimization in scenarios involving small targets and high complexity. Comparative experiment To validate the comprehensive performance of the improved model DAN-YOLO-L, this paper selected multiple mainstream object detection algorithms as comparison models and conducted systematic comparative experiments. The experiments comprehensively evaluated performance across multiple dimensions, including detection accuracy (mAP@0.5, mAP@0.5:0.95), detection reliability (P, R), inference speed (FPS), and computational complexity (GFLOPs). The experimental results are shown in Table 2. To ensure statistical significance and rigorous evaluation, all reported metrics for DAN-YOLO-L and the baseline models represent the average results of three independent training runs with different random seeds. The observed variance for mAP@0.5 across trials was strictly within , thereby confirming that the performance improvements achieved by our method are stable and statistically significant, rather than artifacts of random initialization. The results demonstrate that DAN-YOLO-L achieves the highest detection accuracy. Its mAP@0.5 reached 0.819 and mAP@0.5:0.95 reached 0.514, both surpassing the other comparison models. Specifically, mAP@0.5 improved by approximately 6.4% and 6.1% compared to Faster RCNN (0.770) and SSD (0.772), respectively, significantly enhancing the feature separation capability of traditional detectors in complex backgrounds. Compared to lightweight models like YOLOv5n (0.770), YOLOv6n (0.751), and YOLOv8n (0.797), DAN-YOLO-L achieves superior results on several key metrics while maintaining overall balanced performance., achieving the highest detection accuracy among lightweight models. This demonstrates that structural optimization significantly enhances feature extraction and multi-scale fusion capabilities. Regarding recall (R), DAN-YOLO-L achieves 0.795, representing an approximately 15.6% improvement over YOLOv11n (0.682), thereby achieving higher detection completeness while maintaining high accuracy. Compared to YOLOv10n (0.743) and YOLOv9-t (0.733), recall improved by 7% and 8% respectively, fully demonstrating DAN-YOLO-L’s more robust detection performance for small and weakly textured objects. In terms of inference speed (FPS), DAN-YOLO-L achieves 92.28 FPS, comparable to YOLOv8n (89.53 FPS) and YOLOv9-t (95.0 FPS), while significantly outperforming traditional algorithms like Faster RCNN (25 FPS) and SSD (45 FPS), demonstrating robust real-time detection capabilities. Its computational complexity is 2.84 GFLOPs, slightly lower than YOLOv11n’s 3.2 GFLOPs, and significantly lower than RT-DETR (50.3 GFLOPs) and IRSTD-YOLO (20.0 GFLOPs), achieving an outstanding balance between performance and energy consumption. Notably, the comparison with IRSTD-YOLO33—a dedicated model for infrared small object detection–is particularly striking. DAN-YOLO-L achieves a 3% improvement in mAP@0.5 while significantly reducing computational complexity to approximately one-fifth that of IRSTD-YOLO, with inference speed increasing by about 1.5 times. This demonstrates DAN-YOLO-L’s superior computational efficiency and universal adaptability, particularly in scenarios demanding rapid response and high-density small object detection. Specifically, while IRSTD-YOLO requires 20.04 GFLOPs, our DAN-YOLO-L operates with only 2.84 GFLOPs, representing a reduction in computational complexity by approximately 85.8%. This dramatic improvement in efficiency, achieved without sacrificing detection precision, underscores the superiority of DAN-YOLO-L for deployment on resource-constrained UAV platforms. In summary, DAN-YOLO-L achieves a remarkable equilibrium across detection accuracy, recall, and real-time performance. By introducing multi-scale attention enhancement and lightweight feature fusion mechanisms, it not only effectively improves confidence and robustness in small object detection but also substantially reduces computational resource consumption while maintaining high inference speed. Across all metrics, DAN-YOLO-L demonstrates the optimal performance-efficiency ratio among comparable models, making it particularly suitable for resource- and speed-sensitive application environments such as embedded devices, edge computing terminals, and real-time monitoring scenarios. Table 2Experimental results comparing different algorithms.ModelmAP@0.5mAP@0.5:0.95PRFPS/SGFLOPsFasterRCNN0.7700.5040.7970.67025.005.33SSD0.7720.5050.7900.64045.004.80YOLOv5n0.7700.4900.7980.760109.392.96YOLOv6n0.7510.4920.8120.684102.2011.56YOLOv8n0.7970.5460.8100.71489.5311.79YOLOv9-t0.7820.5050.8330.71095.006.40YOLOv10n0.7670.5050.7840.721103.124.24IRSTD-YOLO0.7950.5080.8400.72060.0020.04RT-DETR0.7900.5010.8550.72355.4050.33YOLOv11n0.7700.5050.8520.68885.193.24DAN-YOLO-L0.8190.5140.8150.79592.282.84Significant values are in [bold]. Experimental results comparing different algorithms. Significant values are in [bold]. To further validate the generalization capability of the proposed model across diverse thermal conditions, we extended our quantitative comparative analysis to the IRay Infrared Dataset. As shown in Table 3, DAN-YOLO-L achieves the highest mAP@0.5 of 0.825, significantly outperforming the baseline YOLOv11n 0.781 and the specialized infrared detector IRSTD-YOLO 0.802. This quantitative evidence confirms that our model effectively generalizes to varied ground-based infrared surveillance scenarios, maintaining superior detection accuracy and efficiency. Table 3Quantitative comparison on the IRay infrared dataset.ModelmAP@0.5PrecisionRecallYOLOv11n0.7810.8200.710IRSTD-YOLO0.8020.8350.750DAN-YOLO-L0.8250.8500.785Significant values are in [bold]. Quantitative comparison on the IRay infrared dataset. Significant values are in [bold]. Ablation experiment To evaluate the impact of different modules on small object detection performance, this study designed ablation experiments to individually validate the contribution of each module. Table 4 presents the results of ablation experiments conducted on the DAN-YOLO model. We performed stepwise optimization based on YOLOv11, incorporating improvements such as the C3K2-D convolution module and auxiliary detection heads, and compared the effectiveness of each approach.The results demonstrate that the optimized model achieves improvements in metrics such as mAP@0.5 and mAP@0.5:0.95. For instance, the baseline YOLOv11 model achieves an mAP@0.5 of 0.770, while incorporating the “C3K2-D + AUX” module increases mAP@0.5 to 0.790. This indicates that adding the convolutional module and replacing the detector head significantly enhances detection accuracy for small objects. Furthermore, adopting the “C3K + AUX + NDW-Inner-CIoU” configuration further elevated mAP@0.5 to 0.799, validating the effectiveness of different module combinations in enhancing feature extraction and fusion. Regarding computational efficiency, the optimized model maintains low computational overhead. For instance, although the introduction of new modules increased GFLOPs, the computational complexity of the model with improved modules (4.03 GFLOPs) did not result in performance degradation compared to traditional YOLOv11 (3.22 GFLOPs). Instead, it enhanced detection accuracy. This demonstrates that our proposed optimization strategy achieves a favorable balance between accuracy and efficiency. Figure 9 shows the detection results of the YOLOv11n model on the Hit-UAV (left) and IRay (right) datasets demonstrating significantly enhanced detection capabilities in complex backgrounds and low-contrast environments compared with the baseline model. Table 4DAN-YOLO ablation experiment.Foundation modelC3K2-DAUXNWD-Inner-CIoUmAP@0.5mAP@0.5:0.95PRGFLOPsParams (M)YOLOv11n0.7700.5050.8520.6883.225.34\checkmark0.7800.5090.8040.7063.165.84\checkmark\checkmark0.7900.4820.8630.6854.036.72\checkmark\checkmark\checkmark0.7990.4990.8510.7054.036.73Significant values are in [bold]. DAN-YOLO ablation experiment. Significant values are in [bold]. Visual comparison of detection results between YOLOv11n and DAN-YOLO-L across diverse typical scenarios. The left panels present sample images from the Hit-UAV dataset, while the right panels display images from the IRay dataset. Comparison of pruning experiments To enhance model inference speed and reduce model size, this study applied L1 pruning optimization to the DAN-YOLO model. It should be emphasized that the L1 pruning optimization was specifically executed on the final, fully-integrated DAN-YOLO model rather than on each intermediate ablation variant. The rationale behind this design is that the primary objective of pruning is to achieve the optimal balance between accuracy and inference efficiency for actual deployment. By performing compression on the highest-performing configuration (which integrates all proposed modules), we ensure that the model retains the maximum representational capacity before undergoing structural regularization. This targeted optimization aligns with the practical goal of providing the most efficient version of our best-performing detector for resource-constrained UAV platforms. L1 pruning evaluates the magnitude of each connection’s weight, removing connections that contribute minimally to the model’s output. This reduces computational load and optimizes the model structure. We performed pruning based on the L1 norm of connections across layers and recovered any potential performance loss through fine-tuning. In experiments, we defined the Speed_up ratio as the computational load of the pruned model divided by that of the original model. By comparing pruning effects at different Speed_up levels, Table 5 shows that when Speed_up reaches 1.6, the pruned model demonstrates significant improvements across multiple performance metrics. Specifically, mAP@0.5 improved from 0.799 to 0.819, while the model’s parameter count and memory footprint decreased by 1.19 M and 2.5 M, respectively. Inference speed also accelerated significantly, validating pruning’s computational efficiency optimization.This slight increase in mAP@0.5 from 0.799 to 0.819 after pruning is primarily attributed to the fine-tuning recovery phase and the structural regularization effect. During fine-tuning, the model re-optimizes the remaining essential weights within a more compact architecture, better capturing the discriminative features of infrared small targets. Additionally, L1-norm pruning filters out redundant channels and parameters that may contribute to background noise or overfitting, leading to a more robust representation. Figure 10 illustrates the comparison of channel counts across layers before and after pruning. It is evident that the channel count in the pruned model is significantly reduced, particularly in high-dimensional layers (e.g., “model_5” to “model_7”), where redundant channels are effectively eliminated. This optimization reduces computational overhead and substantially enhances inference speed. In the figure, yellow bars represent channel counts before pruning, while red bars indicate channel counts after pruning. It is observed that channel counts decreased in most layers, with channel distribution becoming more compact post-pruning, indicating effective model structure refinement. Although mAP@0.5 decreased when the acceleration ratio exceeded 1.6, other performance metrics of the pruned model remained at high levels, particularly inference speed and model size. Therefore, an acceleration ratio of 1.6 was adopted as the final pruning configuration in this study. At this ratio, pruning not only accelerated inference but also effectively reduced model size while maintaining a good balance in accuracy. Table 5Pruning ablation experiment results.Speed_upmAP@0.5mAP@0.5:0.95PRGFLOPsParams(M)1.00.7990.4990.8510.7054.036.731.50.8090.5210.8460.7462.964.781.60.8190.5140.8150.7952.844.231.70.8040.5020.8620.7462.714.011.80.7740.4880.8290.7212.613.751.90.7320.4610.8060.6722.532.882.00.7360.4380.8370.6312.442.53Significant values are in [bold]. Pruning ablation experiment results. Significant values are in [bold]. Channel contrast diagram before and after pruning. The layers from left to right correspond to the Backbone, Neck, and Head modules respectively; yellow bars denote the baseline channels and red bars denote the pruned channels. Figure 10 illustrates the channel distribution from left to right, directly following the architectural sequence defined in Fig. 1. The left portion corresponds to the Backbone (model_0 to model_10), where significant pruning in deep C3K2_DEConv modules removes high-level semantic redundancy. The middle portion represents the Neck (model_11 to model_22), where channel counts in Concat and C3k2 layers remain stable to preserve essential multi-scale feature interactions. The right portion covers the Head section, where the aggressive reduction in auxiliary head channels highlights the elimination of training-only supervisory signals during the inference stage. Regarding the color coding, yellow bars (base) represent the original channel counts in the baseline model, while red bars (prune) indicate the optimized counts retained after L1 structural pruning. The discrepancy between the yellow and red bars visually quantifies the redundant parameters successfully filtered out to improve computational efficiency. Visualization of test results To more intuitively demonstrate DAN-YOLO-L’s advantages over YOLOv11n and IRSTD-YOLO in small object detection, we present detection results from the Hit-UAV and IRay Infrared Dataset using heatmaps. Figure 11 illustrates the performance of DAN-YOLO-L, YOLOv11n, and IRSTD-YOLO across these two datasets. The upper section displays detection results from the Hit-UAV dataset, while the lower section shows results from the IRay Infrared Dataset. The heatmaps reveal that warmer regions indicate areas contributing significantly to the model’s detection results, signifying that objects within these zones played a more critical role in the detection process. We confined the heatmaps to the detection bounding boxes to more precisely observe the confidence distribution for each target. On the Hit-UAV dataset, DAN-YOLO-L demonstrated higher small object detection accuracy compared to YOLOv11n and IRSTD-YOLO, with a more uniform confidence distribution across target regions and a significant reduction in missed detections. Particularly in complex backgrounds and areas densely populated with small targets, DAN-YOLO-L effectively identifies small objects. In contrast, YOLOv11n and IRSTD-YOLO exhibit scattered bounding box distributions and lower confidence levels, resulting in undetected small targets. On the IRay Infrared Dataset, DAN-YOLO-L also performed exceptionally well, particularly in detecting small objects at long distances and under low-contrast conditions. Heatmaps reveal that DAN-YOLO-L generates more concentrated and precise bounding boxes, whereas YOLOv11n and IRSTD-YOLO produce more scattered detection boxes with lower confidence scores, resulting in the failure to identify some small objects.In addition, to provide a more intuitive comparison of the evaluation metrics between YOLOv11n and the improved algorithm, Fig. 12 illustrates the contrasts in mAP@0.5, precision, and recall, clearly demonstrating the superior performance of the proposed DAN-YOLO-L model. Comparison of thermal maps at different scales: Left: Results on the Hit-UAV dataset Right: Results on the IRay Infrared Dataset (a) YOLOv11n (b) IRSTD-YOLO (c) DAN-YOLO-L. Performance metrics comparison chart. Conclusions This paper addresses issues such as feature loss, inaccurate localization, and limited real-time performance caused by scale differences in small object detection by proposing the improved model DAN-YOLO. The model integrates deconv and C3k2 modules to enhance multi-scale feature extraction, introduces an improved Auxiliary Head for efficient multi-layer feature fusion, and designs an NWD-Inner-CIoU loss function combining Wasserstein distance with an inner auxiliary box mechanism to boost small object localization accuracy and robustness. Additionally, the proposed L1 pruning strategy effectively reduces computational complexity, improving inference speed while maintaining accuracy. Experiments on the Hit-UAV and IRay Infrared Dataset demonstrate that DAN-YOLO outperforms YOLOv11n in metrics including mAP@0.5, mAP@0.5:0.95, Precision, and Recall, particularly excelling in small object detection within complex backgrounds and low-contrast scenarios. The model maintains high efficiency in terms of GFLOPs and FPS, validating its potential for real-time applications in drones and embedded devices. Future work will focus on further enhancing the model’s generalization capabilities and lightweight design to accommodate more complex environments and multi-scenario tasks. Author contributions K.F. Designed and performed the experiments, analyzed the data, and prepared the manuscript. G.J.L. and T.L. Contributed to methodology refinement and manuscript revision. Z.Q.L. Assisted in data review and discussions during manuscript preparation. R.M. Participated in discussions regarding the experimental design and provided valuable insights for data analysis and manuscript revisions. All authors approved the final manuscript. Funding This work was supported by the National Natural Science Foundation of China (Grant No. 42405145); the Sichuan Provincial Science and Technology Department Project (2024YFHZ0026); and the Postgraduate Course Case Construction Project of Sichuan University of Science and Engineering (AL202306). Data availability The datasets generated and/or analyzed during this study are available at the following links: The HIT-UAV dataset is publicly available at https://github.com/suojiashun/HIT-UAV-Infrared-Thermal-Dataset. The iRay Optoelectronics infrared open-source datasets are available from the official infrared open platform of iRay Optoelectronics at https://openai.iraytek.com/apply/Other_plate.html. Researchers can access the datasets for academic and non-commercial use according to the respective data usage terms. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and/or analyzed during this study are available at the following links: The HIT-UAV dataset is publicly available at https://github.com/suojiashun/HIT-UAV-Infrared-Thermal-Dataset. The iRay Optoelectronics infrared open-source datasets are available from the official infrared open platform of iRay Optoelectronics at https://openai.iraytek.com/apply/Other_plate.html. Researchers can access the datasets for academic and non-commercial use according to the respective data usage terms.