Optimized YOLO architectures for efficient Kiwi detection in precision agriculture on embedded systems Authors: Berat Karacaoglu, Muhammet Emin Sahin Published: 2025 ABSTRACT AbstractAccurate and efficient fruit detection is essential for precision agriculture, particularly in densely occluded crops such as kiwifruit. This study presents a comprehensive benchmarking and optimization framework covering YOLOv8–YOLOv11 architectures for kiwi detection, evaluated under both high-performance training conditions and embedded deployment on an NVIDIA Jetson TX2. A field-collected dataset containing 2,925 training and 1,936 test annotations was used to train five sub-models (n, s, m, l, x) per YOLO version under identical settings. To enhance efficiency for edge deployment, a structured hyperparameter optimization procedure was applied to all “s” models, yielding substantial performance gains without additional architectural modifications. Among all evaluated models, the optimized YOLOv11s achieved the best accuracy–efficiency trade-off, reaching mAP@0.5 = 0.956, precision = 0.868, recall = 0.918, and an embedded inference time of 3.33 s/image on Jetson TX2. While larger models (e.g., YOLOv8x, YOLOv11l) attained slightly higher raw accuracies (up to mAP@0.5 = 0.957), their latency rendered them unsuitable for edge deployment. The results demonstrate that lightweight YOLO architectures, when supported by targeted hyperparameter tuning, can be effectively adapted for resource-constrained agricultural systems. The proposed evaluation and optimization pipeline provides a transferable methodology for other fruit-detection tasks and supports future development of embedded vision solutions in precision agriculture.Keywords: Kiwi detection, YOLO, Object detection, Embedded systems, Precision agricultureSubject terms: Computational biology and bioinformatics, Mathematics and computing, Plant sciences Introduction Agriculture, one of humanity’s oldest economic pillars, remains vital for food security, ecological sustainability, and socioeconomic stability1,2. Despite technological progress, agriculture faces unprecedented challenges: rapid population growth increases food demand, while arable land continues to shrink due to urbanization and climate change2. Environmental stressors—such as drought, floods, extreme temperatures, and water scarcity—further disrupt production processes and expose the limitations of traditional farming practices3. These pressures have accelerated the transition toward smart, data-driven agriculture, where productivity is evaluated not only by yield but also by resource efficiency, environmental impact, product quality, and labor optimization4. Within this transformation, digitalization, automation, and artificial intelligence (AI) have become central. Especially in fruit production—where processes such as harvest timing, quality assessment, ripeness estimation, and early disease detection still rely heavily on human expertise—AI-driven systems offer crucial gains in speed and accuracy2. Manual observation is time-consuming, error-prone, and insufficient for the early detection of diseases or optimal harvest decisions, which may lead to reduced shelf life, lower quality, and economic loss. A core challenge in these environments is the reliable detection of fruits under natural conditions where foliage, occlusion, shadows, and varying illumination significantly complicate visual recognition. Kiwi fruit exemplifies these difficulties due to its dense canopy structure, overlapping growth patterns, and skin color that blends with the background5. Deep learning–based computer vision methods, particularly convolutional neural networks (CNNs), have demonstrated strong performance in overcoming these challenges6. YOLO (You Only Look Once)–based detectors are particularly suited for real-time agricultural applications because they can localize multiple objects in a single inference pass7. Several studies confirm the robustness of YOLO variants for fruit detection, ripeness analysis, and yield estimation in complex outdoor environments. For instance, a YOLOv7-based plum detection system achieved high accuracy under occlusion and variable lighting9, while Chen et al. enhanced YOLOv8n with EMBC, DySample, and TADDH modules, producing EDT-YOLOv8n, which improved accuracy and reduced model size for embedded systems8. Similarly, Niu et al. redesigned YOLOv5 for small unripe plum detection, reaching 91.65% accuracy with only 7 ms inference time, outperforming YOLOv5, YOLOv7, and Faster R-CNN9. Studies across various crops highlight the comparative advantages of YOLO-based architectures. Li et al. presented YOLOv3 to provide the best balance of speed and accuracy for greenhouse detection tasks compared with Faster R-CNN and SSD10. An et al. proposed SDNet (based on YOLOX), achieving 94.26% precision by integrating attention mechanisms and enhanced bounding-box optimization11. Xu12 applied the YOLOv3 detection algorithm to identify and classify tea buds across different growth stages and categories. Lightweight and attention-enhanced YOLO variants continue to expand capabilities for detecting small, overlapping fruits. Examples include GhostNet-YOLOv4 (Gao et al.) and YOLOv8n-CRS (Chen et al., 2024), both of which improve performance while reducing computational cost13,14. Tang et al. submitted 94.91% mAP with YOLOv7-plum by integrating CBAM attention15. Beyond fruit detection, YOLO-based systems facilitate wildfire monitoring16, plant disease recognition17, and other agricultural decision-support tasks. Recent advancements in optimization and hybrid AI contribute to performance improvements across domains relevant to smart agriculture. Ghosh et al. proposed the MLOA algorithm, which overcame early-convergence issues and achieved high accuracy in medical image analysis18, while Chanda and Ghosh demonstrated how AI–blockchain integration can streamline HR processes19. Transformer–CNN hybrid models have also shown promise in physics-guided image analysis, reducing turbulence prediction errors by up to 40%20. Although these studies are outside agriculture, they exemplify methodological innovations—metaheuristics, hybrid architectures, and efficient optimization—that can inform future agricultural vision systems. Parallel to detection research, IoT-based smart agriculture systems are evolving rapidly. Embedded irrigation systems using real-time environmental data significantly reduce water waste, as demonstrated by Morchid et al.21. Fuzzy logic and IoT-based irrigation architectures have been shown to decrease water consumption by 11–28% and increase yields by up to 37%22. Drone technologies, highlighted by Chouhan et al., submitted fast, accurate data for disease detection, spraying, and growth monitoring, contributing to precision agriculture and sustainable farm management23,24. Chouhan et al. developed a Counter Propagation Network trained with Fuzzy Competitive Learning (FCPN) to handle uncertainties in natural scene images25. Using region growing for initial seed generation and Bacterial Foraging Optimization for weight tuning, the model automatically determined the optimal number of clusters and achieved high object-segmentation accuracy, outperforming FCM, SOM, CPN, and SEEFC. Upadhyay et al. proposed 3SW-Net, a lightweight architecture for early weed detection combining grayscale imagery, SLIC superpixel segmentation, and HOG-based texture features26. Evaluated on 4,000 images using five-fold cross-validation, it achieved 99.56% accuracy, 98.99% sensitivity, and 99.68% specificity. With only ~ 9 million parameters, 3SW-Net is well suited for embedded and real-time smart agriculture systems. Upadhyay and Gupta conducted a series of studies focusing on plant disease detection and severity assessment across multiple fruit species. In their first work, they introduced a modified ResNeXt-based framework that combines preprocessing, Fuzzy C-Means segmentation, and an enhanced ResNeXt classifier for identifying fungal diseases in heterogeneous fruit images, including apples, guavas, and custard apples. Evaluated on 14,408 images, the model achieved 98.92% accuracy and F1-scores exceeding 0.99, outperforming existing multi-crop disease detection methods27. In a subsequent study, the authors proposed a disease-severity assessment system that integrates a ResNet-50 classifier with SegLearner—an improved SegFormer-derived segmentation architecture featuring multi-scale feature aggregation, an efficient Transformer encoder, and semi-supervised mask generation. Tested on the Mask-LeafDataset, SegLearner achieved 91.4% mIoU, 95.4% accuracy, and a 94.7% F1-score, outperforming U-Net, SegFormer, and Mask R-CNN. By leveraging pixel-ratio calculations from segmentation masks, the system demonstrated reliable severity categorization and strong generalization on an external dataset28. Paul and Machavaram evaluated YOLOv9c-seg and YOLOv9e-seg for peduncle detection in nighttime greenhouse images, where YOLOv9c-seg delivered the best results with mAP@0.5 scores of 0.751 (boxes) and 0.725 (masks) and an F1-score of 0.89; Grounded SAM, operating without training, achieved only 59% confidence29. Their automated harvesting framework, based on YOLOv8s, achieved mAP@0.5 = 0.967 for fruit detection, box mAP = 0.790, and mask mAP = 0.771. Using a RealSense D455 camera, 3D cut-point localization errors were 8–12 mm, while the ByteTrack counting module reached 94.1% accuracy and the mobile app delivered real-time detection with IoU = 0.92 and 150 ms latency30. In a subsequent nighttime detection study, YOLOv9 again surpassed Grounding DINO with mAP = 0.947 and 26 ms inference, compared with Grounding DINO’s F1-score of 0.70. Yet, Grounding DINO showed greater flexibility in zero-shot tasks such as positional search and growth-stage estimation. Leveraging both strengths, the authors introduced YOLOVOVOD, which increased Grounding DINO’s positional-search confidence by 65.11% and achieved 88% confidence in growth-stage detection—offering a more accurate and adaptable framework for agricultural machine vision31. Deep learning remains central in plant disease diagnosis, with numerous studies reporting near-perfect accuracies. Pacal et al. analyzed 160 studies from 2020 to 2024, demonstrating that CNN and Vision Transformer (ViT) models provide reliable automated disease detection32. EfficientNet, ResNet, Res2Net, and MaxViT-derived architectures have consistently achieved high accuracies in sugarcane33, tomato34, grapevine35, and apple36, disease classification37,38. These results support the broader argument that advanced CNN/ViT architectures significantly enhance agricultural decision-making. Other supporting studies include weed detection improvements via PSO-based YOLO optimization39, blockchain-based supply chain solutions40, cloud-enabled precision agriculture platforms41, and enhanced classification ensemble models for disease detection42. Furthermore, several works focusing on capsicum detection—YOLOv843, and YOLOv10 vs. SAM-DINO44 highlight the growing applicability of YOLO for real-time greenhouse monitoring. Beyond agriculture-specific applications, broader advances in digital ecosystems also inform the development of smart farming technologies. For example, Chanda et al. introduced a blockchain-based offboarding framework using the PBFT consensus mechanism, demonstrating secure, transparent, and high-throughput event management45. Similarly, Chanda and Singh conducted a bibliometric analysis of blockchain use in HR management, reviewing 124 studies and identifying key themes such as recruitment, talent governance, and data security46. Morchid et al. show that AI- and IoT-based systems improve early detection of agricultural and forest fires, increasing response speed and resource efficiency, and achieving higher accuracy than traditional methods47 Paul and Machavaram report that GenAI enhances efficiency, decision-making, and resource use in Indian agriculture within the Agriculture 4.0 framework, though infrastructural, data-governance, and ethical issues still limit broader adoption. Their study presents a feasible framework integrating multiple data sources with GenAI, supported by pilot project outcomes48. Although these works lie outside the agricultural domain, they highlight emerging trends in decentralized data management and trustworthy AI systems—concepts that increasingly shape next-generation smart agriculture infrastructures. In the context of a public kiwi fruit dataset used by Ulutaş and Altın49, Mask R-CNN with a ResNet50 backbone achieved a mAP of 98.48%, demonstrating the feasibility of deep learning–based kiwi detection despite occlusion and lighting variations. In this study, we present a comprehensive evaluation of the YOLOv8–YOLOv11 models for kiwi detection by integrating hyperparameter optimization and embedded deployment on the NVIDIA Jetson TX2 platform, with the goal of identifying the most efficient detection configuration capable of achieving near-real-time performance under embedded system constraints. An overview of study is shown in Fig. 1. The main contributions of this study can be summarized as follows: An overview of study. A systematic comparison of YOLOv8–YOLOv11 models for kiwi detection in natural orchard environments, covering both high-performance and embedded deployment scenarios. A systematic comparison of YOLOv8–YOLOv11 models for kiwi detection in natural orchard environments, covering both high-performance and embedded deployment scenarios. Targeted optimization of the “s” variants (YOLOv8s, YOLOv9s, YOLOv10s, YOLOv11s), leading to improved accuracy, stability, and inference efficiency without additional computational overhead. Targeted optimization of the “s” variants (YOLOv8s, YOLOv9s, YOLOv10s, YOLOv11s), leading to improved accuracy, stability, and inference efficiency without additional computational overhead. Implementation and benchmarking of all models on the NVIDIA Jetson TX2 platform, providing practical insights into near-real-time performance in resource-constrained environments. Implementation and benchmarking of all models on the NVIDIA Jetson TX2 platform, providing practical insights into near-real-time performance in resource-constrained environments. Demonstration that YOLOv11s, when optimized, offers the best trade-off between accuracy (mAP@0.5 of 0.956) and inference time (3.33 s), making it the most suitable model for real-world kiwi detection. Demonstration that YOLOv11s, when optimized, offers the best trade-off between accuracy (mAP@0.5 of 0.956) and inference time (3.33 s), making it the most suitable model for real-world kiwi detection. Providing a transferable methodology and performance analysis framework that can be applied to other fruit detection tasks in similar complex agricultural conditions. Providing a transferable methodology and performance analysis framework that can be applied to other fruit detection tasks in similar complex agricultural conditions. The structure of this paper is organized as follows: Sect. 2 describes the materials and methods, including dataset preparation, pre-processing, YOLO configurations, and hyperparameter optimization. Section 3 presents the experimental setup, detailing the computational environment for both high-performance training and embedded deployment. Section 4 reports a discussion of the findings, focusing on the trade-offs between accuracy and inference speed in real-world applications. Finally, Sect. 5 concludes the study, summarizing the main contributions and highlighting the potential of the proposed approach for broader precision agriculture applications. Materials and methods Dataset The image dataset used in this study was collected in August 2021 at Güngör Farm, located in the Çarşamba Plain of Samsun, Turkey, during the pre-harvest period of kiwifruit under natural orchard conditions49. To ensure visual diversity and to reflect real-world agricultural scenarios, all images were captured on-site under naturally varying illumination, shadowing, canopy density, and branch complexity. Unlike laboratory-prepared datasets that often present controlled and homogeneous imaging conditions, this dataset incorporates the environmental variability inherent in open-field production systems, thereby increasing the model’s robustness and practical applicability. Image acquisition was carried out using a Samsung A51 smartphone mounted on a tripod-supported gimbal system fixed at approximately 1 m above ground level. The device’s 48-megapixel RGB main camera enabled the collection of high-resolution images under natural daylight without any artificial lighting. This setup was intentionally chosen to capture realistic orchard scenes where fruit visibility is affected by leaf occlusions, irregular branch structures, and heterogeneous light distribution. Such natural variability was included to enhance the dataset’s representativeness and improve the model’s ability to generalize under real operating conditions. The dataset used in this study has been made publicly accessible on GitHub, allowing other researchers to utilize it for benchmarking and reproducibility purposes50. For annotation, rectangular bounding boxes were manually created exclusively for object detection tasks. Since the dataset contains only a single class (kiwifruit), all annotations were assigned the class label “0.” The dataset is divided into two subsets: the training set consists of 66 images containing a total of 2,925 bounding box annotations, while the test set includes 40 images with 1,936 bounding box annotations. Representative raw images from the dataset are illustrated in Fig. 2. These images were captured directly in natural orchard conditions, exhibiting significant visual variability such as heterogeneous illumination, leaf density, branch occlusions, and irregular shadow patterns. This natural diversity creates a realistic and challenging learning environment that enhances the practical applicability of object detection models. Representative examples from the dataset: Manual annotation was performed using the Roboflow platform, where each kiwifruit instance was carefully enclosed within a rectangular bounding box. An example of the annotation interface is shown in Fig. 3. The inclusion of accurate and detailed bounding boxes ensured high-quality supervision during model training. In deep learning–based object detection, particularly in single-stage architectures such as YOLO, model performance strongly depends on the quality and consistency of input images. Therefore, raw images underwent several preprocessing steps before being fed into the model to improve accuracy and stabilize training. All images, originally captured at 512 × 512 pixels, were resized to 640 × 640 pixels using the Roboflow platform to match the input size requirements of the YOLO architecture. EXIF metadata was used to automatically correct image orientation, ensuring consistent object alignment across the dataset. Manual bounding boxes created using the Roboflow platform. The YOLO model’s fixed input resolution facilitated standardized training and testing, enabling improved learning of objects at varying scales. Bounding box coordinates were normalized relative to image dimensions and stored in YOLO format as [x_center, y_center, width, height] along with class labels Since the dataset contains only a single class (kiwifruit), all annotations were assigned the class label “0.” During the annotation process, bounding boxes were drawn to tightly enclose each visible kiwifruit instance, regardless of variations in size, occlusion, or viewing angle. In cases where fruits overlapped, each fruit was annotated individually as long as its contour was partially identifiable. For partially occluded fruits, bounding boxes were created if at least 30–40% of the fruit surface was visible. Small and distant fruits were also annotated, provided that their circular shape remained distinguishable from the background. This consistent annotation policy ensured uniform bounding box placement across the dataset, particularly in visually complex orchard scenes with dense foliage and irregular branch structures. These preprocessing steps ensured more stable training and enhanced the model’s robustness against environmental variability commonly present in orchard conditions. Object detection and YOLO models Object detection, one of the most impactful applications of deep learning, aims not only to identify objects within an image but also to determine their precise locations51. Traditional image processing methods have shown limited capability in this domain, but the advent of Convolutional Neural Networks (CNNs) brought a major breakthrough. In challenging scenarios with cluttered backgrounds or overlapping objects, deep learning models stand out by performing classification and localization simultaneously52. Region proposal–based methods, such as Faster R-CNN, achieve high accuracy but often struggle to meet the speed requirements of real-time applications10. Single-stage architectures, developed to optimize both speed and accuracy, address these limitations. Among them, YOLO introduced a paradigm shift by performing detection through a single neural network, delivering an effective balance between accuracy and real-time performance7. The evolution of the YOLO architecture began with YOLOv1, and each subsequent release has introduced more advanced network structures, improved training strategies, and higher detection performance. The early versions, from YOLOv1 to YOLOv3, primarily focused on speed and architectural simplicity. The intermediate versions, YOLOv4 through YOLOv7, integrated enhanced modules to optimize the trade-off between accuracy and computational efficiency. With the release of YOLOv8 in 2023, the architecture adopted an anchor-free design, improving flexibility and reducing computational complexity. Subsequent versions—YOLOv9, YOLOv10, and YOLOv11—refined the architecture with advanced building blocks, gradient optimization techniques, and attention mechanisms. These updates led to notable improvements in both small-object detection and overall accuracy. This section presents a chronological overview of YOLO’s evolution, highlighting the architectural innovations and their contributions to detection performance across versions. Table 1 presents a comparative overview of the YOLOv8–YOLOv11 models employed in this study, highlighting their architectural differences and core technical characteristics. For each model, the backbone structure, primary architectural blocks, head design, and key algorithmic innovations are systematically summarized. This comparison illustrates the evolutionary trajectory of the YOLO family—from anchor-free detector designs to NMS-free paradigms and, most recently, to hybrid CNN–Transformer architectures. It clarifies the rationale behind selecting these models with respect to their strengths in detection accuracy, computational efficiency, inference speed, and small-object recognition capabilities. By providing a transparent summary of the architectural components used, the table strengthens the reproducibility and methodological clarity of the experimental framework adopted in this study. Table 1Architectural comparison of YOLOv8–YOLOv11.ModelBackboneCore Architectural BlocksHeadKey InnovationsAdvantagesYOLOv8C2f BackboneC2f blocks, SPPF, PANDecoupled HeadAnchor-free design, multi-task capabilityFaster, lighter, mobile-friendlyYOLOv9GELANGELAN structure, RepConv variantsStandard YOLO HeadPGI (Programmable Gradient Information), enhanced feature fusionStrong small-object performance, more stable gradient flowYOLOv10Fast–Slow architectureSCDD (Spatial–Channel Decoupled Downsampling), LKCDual Head: One-to-One & One-to-ManyNMS-free detection, fully differentiable pipelineLess information loss, high FPSYOLOv11Hybrid CNN–TransformerC3k2, C2PSA attentionEnhanced decoupled headHybrid CNN–Transformer framework with attention-based feature processingSuperior performance on overlapping objects Architectural comparison of YOLOv8–YOLOv11. Hyperparameters Hyperparameters are user-defined variables that configure a machine learning algorithm prior to training and determine how it operates. Unlike model parameters, which are learned during training, hyperparameters are set in advance and exert a decisive influence on model accuracy, convergence speed, and generalization capacity. To identify the most appropriate configuration, the literature proposes various hyperparameter optimization strategies. One of the most widely used methods is Grid Search (GS), which systematically evaluates all possible combinations within predefined value ranges for each hyperparameter. Although GS is straightforward and easy to implement, its computational cost increases substantially as the number of hyperparameters grows, due to the exponential expansion of the search space. Consequently, more efficient alternatives—such as Random Search, Bayesian Optimization, and evolutionary algorithms—are increasingly preferred for high-dimensional problems. Random Search mitigates unnecessary evaluations by randomly sampling the search space and can outperform GS, particularly when certain hyperparameters have disproportionately greater influence on performance. Bayesian Optimization, on the other hand, leverages knowledge from prior evaluations to guide the search process, enabling near-optimal solutions with fewer iterations. Evolutionary algorithms draw on principles of biological evolution to offer strong exploratory capabilities within large and complex search spaces. The common advantage of these methods is their ability to reduce computational cost while improving model performance through more intelligent and adaptive search mechanisms. In this context, modern hyperparameter optimization approaches provide considerably more practical, effective, and sustainable solutions for high-dimensional and computationally expensive deep learning models compared with traditional techniques. Evaluation metrics In image-based object detection, the performance of deep learning architectures is assessed not only in terms of theoretical accuracy but also by their stability in real-world applications. This is particularly crucial in domains with low error tolerance—such as agriculture, healthcare, security, and industrial automation—where models must demonstrate not only the ability to make correct detections but also the capability to produce consistent and reliable results. Real-time object detection frameworks, such as YOLO, offer significant advantages in this regard7,53. To objectively evaluate and compare the performance of such algorithms, systematic evaluation metrics are essential. In object detection tasks, the most commonly adopted metrics include Intersection over Union (IoU), Precision, Recall, F1 Score, and mean Average Precision (mAP)54–57. These metrics provide quantitative insights into both the localization accuracy and the detection coverage of the model. Intersection over Union (IoU) is a fundamental accuracy metric that measures the spatial overlap between the predicted object location and the ground truth bounding box. It quantifies how precisely the predicted bounding box aligns with the actual object54. In YOLO architectures, the input image is divided into grid cells, each responsible for predicting bounding boxes. The accuracy of these predictions is evaluated using the IoU score. During training, predictions with low IoU values are penalized by the loss function, thereby improving the model’s localization precision. Additionally, IoU serves as a threshold criterion in the calculation of higher-level metrics such as mAP. The formulas given in Eqs. 1–5: where represents the predicted bounding box, and denotes the ground truth bounding box. A higher IoU indicates better alignment between the prediction and the ground truth. In the context of kiwifruit detection, IoU is particularly significant because kiwifruits often appear partially occluded by leaves or clustered together in orchard environments; therefore, higher IoU values directly reflect the model’s ability to localize fruits accurately even under occlusion and overlapping conditions. Precision reflects the proportion of correct positive detections among all predicted positives: where TP denotes True Positives and FP denotes False Positives. In YOLO, high precision ensures that most detected bounding boxes correspond to actual objects, which is critical in applications where false alarms are costly. For kiwifruit detection, high Precision is crucial as the visual similarity between kiwi surfaces, leaves, and shaded background regions can easily lead to false positives; thus, higher Precision indicates the model’s robustness against such misleading patterns. Recall measures the proportion of correctly detected objects among all actual objects: where FN denotes False Negatives. High recall is essential in YOLO for minimizing missed detections, especially in applications like agriculture and safety monitoring. In orchard conditions where kiwifruits frequently appear in shadowed areas, behind branches, or in overlapping clusters, high Recall ensures that all visible fruits are captured, minimizing missed detections that are critical for yield estimation and automated harvesting. mAP is the most widely adopted metric for object detection benchmarks. It calculates the mean of Average Precision (AP) values across all classes and IoU thresholds. For a single class, AP is computed as the area under the Precision–Recall curve and mAP is given below: where denotes the number of classes. In YOLO, mAP@0.5 uses an IoU threshold of 0.5, while mAP@0.5:0.95 averages results across thresholds from 0.5 to 0.95 in 0.05 increments. In this study, mAP@0.5 is adopted as the primary evaluation metric due to its stable, interpretable, and widely accepted use in agricultural object detection. Although mAP@0.5:0.95 values are also reported for completeness, the main analyses and comparisons are based on mAP@0.5. Embedded system In this study, the NVIDIA Jetson TX2 was employed as a high-performance edge computing platform optimized for embedded artificial intelligence applications58. Equipped with a 256-core Pascal-based GPU and a heterogeneous six-core CPU architecture consisting of dual Denver 2 and quad ARM Cortex-A57 cores, the Jetson TX2 provides low-latency and energy-efficient processing for real-time image analysis and deep learning inference. Through the JetPack SDK, the platform offers CUDA, cuDNN, TensorRT, and OpenCV support, enabling hardware-accelerated and highly optimized execution of deep learning models. The key hardware and software components of the Jetson TX2 used in this study are summarized in Table 2. Table 2Jetson TX2 hardware and software Components.CategoryFeatureDescription / ValueGPUNVIDIA Pascal GPU256 CUDA coresCPU6-core CPU2× Denver 2 + 4× ARM Cortex-A57MemoryRAM8 GB LPDDR4StorageInternal storage32 GB eMMCCompute PowerFP16 performance~ 1.33 TFLOPSPower ConsumptionTypical operation7.5–15 WConnectivityI/OUSB 3.0, HDMI, MIPI CSI, PCIeSoftware SupportJetPack SDKCUDA, cuDNN, TensorRT, OpenCVOperating SystemUbuntu-basedLinux for Tegra (L4T)Application AreasEdge AI, RoboticsObject detection, image processing, autonomous systems Jetson TX2 hardware and software Components. Jetson TX2’s GPU architecture, CPU configuration, memory capacity, power profile, and extensive software ecosystem collectively demonstrate why the device is well suited for edge AI applications. This integrated hardware–software architecture enables stable, efficient, and real-time execution of deep learning tasks—such as object detection, image processing, and autonomous system control—directly on the device with minimal latency58. Consequently, the Jetson TX2 provides an effective and reliable platform for implementing the computationally intensive AI workloads required in this study. Experimental results This section presents the experimental findings from the training and testing of the AI-based system developed for the real-time detection of kiwi fruit. The study evaluates different versions of the YOLO algorithm in terms of training configurations, performance analysis, hyperparameter optimization, and real-time deployment on embedded systems. The goal is to identify the most effective architectures for different operational scenarios by comparing models in terms of accuracy, recall, and inference time. In the experimental process, four versions of the YOLO algorithm—YOLOv8, YOLOv9, YOLOv10, and YOLOv11—were employed for kiwi detection. To ensure a fair comparison, a consistent and standardized training configuration was applied to all versions. This approach allowed architectural differences to be isolated as the primary factor influencing performance, while variability caused by training conditions was minimized. The key parameters used during the training process are summarized in Table 3. Table 3Key training parameters.ParameterValueDescriptionepochs250 (max)Maximum training cycles; early stopping enabled.patience50Training stops if no improvement is observed in validation performance.batch8Number of images per iteration.imgsz640Input image resolution (pixels).optimizerAdamWOptimization algorithm for weight updates.lr00.0005Initial learning rate.workers4Number of threads for data loading during training.device0GPU device identifier (CUDA:0).single_clsTrueSingle-class (kiwi) object detection. Key training parameters. The model training and evaluation processes are conducted on a high-performance computing system to efficiently handle the training of computationally intensive architectures and to test real-time detection scenarios. The experimental setup is equipped with 64 GB of RAM and dual Intel® Xeon® Silver 4114 CPUs operating at 2.20 GHz with 20 cores in total. The system utilized an NVIDIA GeForce RTX 3090 Ti GPU to accelerate deep learning computations. All experiments were performed on an HP Z6 G4 workstation running Windows 10. The implementation is developed in Python 3 within the Visual Studio Code environment. YOLO models YOLOv8 submodels: overview and performance evaluation The performance analysis of the YOLOv8 submodels is carried out using the best.pt weights obtained at the point of highest validation performance for each model. The YOLOv8 architecture offers five scalable submodels—n, s, m, l, and x—designed to accommodate different hardware capabilities and application requirements. These models vary in terms of parameter count, computational complexity, and processing load, enabling deployment across a broad spectrum of systems, from low-resource edge devices to high-performance computing platforms. The comparative performance metrics of the YOLOv8 submodels in kiwi fruit detection are summarized in Table 4. Table 4Performance metrics of YOLOv8 submodels for Kiwi fruit detection.ModelIterationsPrecisionRecallF1 ScoremAP@0.5mAP@0.5:0.95YOLOv8n920.8850.8910.8880.9510.620YOLOv8s920.8750.9040.8890.9530.627YOLOv8m560.8790.8730.8760.9420.628YOLOv8l610.8740.8750.8750.9460.636 YOLOv8x 56 0.883 0.899 0.891 0.956 0.640 Performance metrics of YOLOv8 submodels for Kiwi fruit detection. The results indicate that YOLOv8 models deliver high detection performance, with each variant exhibiting characteristic strengths across the evaluated metrics. The YOLOv8x model achieved the highest overall accuracy, obtaining an F1 score of 0.891, mAP@0.5 of 0.956, and mAP@0.5:0.95 of 0.640, making it the most reliable architecture in terms of comprehensive detection capability. Among the lightweight variants, YOLOv8s recorded strong precision (0.875) and the highest recall (0.904) within the family, effectively reducing missed detections while maintaining computational efficiency. Although the YOLOv8n model produced slightly lower scores, its fast inference and minimal resource usage make it suitable for real-time, embedded deployments. Medium-scale models such as YOLOv8m and YOLOv8l provided balanced results, offering incremental gains in mAP values at the expense of higher computational load. Regarding training duration, the YOLOv8n and YOLOv8s models converged more rapidly than the larger variants, representing the most efficient options in terms of training speed. In contrast, the YOLOv8l and YOLOv8x models required more iterations to achieve stable performance due to their increased parameter counts and architectural complexity. Despite this longer optimization process, their higher accuracy makes them preferable for high-precision detection systems without strict hardware limitations. Supporting this evaluation, performance curves and iteration-based trends for YOLOv8s and YOLOv8x are presented in Figs. 4 and 5. YOLOv8s: a Model performance evaluation metrics, b Training and validation loss curves. YOLOv8x: a Model performance evaluation metrics, b Training and validation loss curves. Figures 4 and 5 illustrate the evolution of the three primary loss components—box loss, classification loss, and distributional focal loss—throughout the training of the YOLOv8s and YOLOv8x models. In both architectures, the loss values start at relatively high levels and progressively decrease, confirming stable convergence. The YOLOv8s model demonstrates a smooth decline in training losses with validation curves closely tracking the same trend, indicating strong generalization and limited overfitting. The YOLOv8x model exhibits a similar pattern, with validation losses remaining consistently aligned with training losses despite the model’s larger capacity. This stable relationship across epochs suggests that the YOLOv8x architecture effectively utilizes its higher complexity without exhibiting divergence or overfitting, reinforcing its role as the most accurate variant in the YOLOv8 family. YOLOv9 submodels: overview and performance evaluation The performance analysis of the YOLOv9 submodels is conducted using the best.pt weights obtained at the point of highest validation performance for each model. The YOLOv9 family includes five variants—t, s, m, c, and e—that differ in architectural complexity, parameter density, and computational requirements. This diversity enables flexible model selection tailored to different system constraints. The comparative performance metrics of the YOLOv9 submodels for kiwi fruit detection are presented in Table 5. Table 5Performance metrics of YOLOv9 submodels for Kiwi fruit detection.ModelIterationsPrecisionRecallF1 ScoremAP@0.5mAP@0.5:0.95YOLOv9t810.8760.8850.8800.9480.623YOLOv9s680.8640.8950.8790.9460.630YOLOv9m620.9000.8700.8850.9500.635YOLOv9c840.8800.8780.8790.9410.633 YOLOv9e 66 0.887 0.890 0.889 0.951 0.645 Performance metrics of YOLOv9 submodels for Kiwi fruit detection. The results demonstrate that YOLOv9 models generally achieve high accuracy, with each variant displaying distinct detection characteristics. The highest mAP@0.5 value 0.951 is obtained with YOLOv9e, which also achieved an F1 score of 0.889, making it the top-performing model in overall detection capability. YOLOv9m recorded the highest precision 0.900, effectively minimizing false positives. In terms of recall, YOLOv9s achieved the highest value 0.895, offering an advantage in reducing missed detections. The YOLOv9c and YOLOv9t models produced balanced results; however, their performance in more stringent metrics such as mAP@0.5:0.95 was slightly lower compared to the top-performing variants. Regarding training efficiency, YOLOv9m reached optimal performance in just 62 iterations, marking the shortest convergence time among all submodels, while YOLOv9c required 84 iterations, the longest training duration. This disparity underscores the importance of training duration as a crucial factor in model selection. Consequently, lightweight models such as YOLOv9s and YOLOv9t are well-suited for resource-constrained real-time applications, whereas YOLOv9e and YOLOv9m are preferable for high-accuracy systems without strict hardware limitations. These training dynamics are supported by the iteration-based loss curves and performance trends for YOLOv9m and YOLOv9e, which are illustrated in Figs. 6. and 7. respectively. YOLOv9m: a Model performance evaluation metrics, b Training and validation loss curves. Figures 6 and 7 show the evolution of the three primary loss functions—box loss, classification loss, and distributional focal loss—during the training process. Both models exhibit high initial loss values that steadily decrease over time, confirming successful convergence. Regarding generalization, the YOLOv9m model demonstrates a close alignment between training and validation losses, suggesting minimal overfitting and strong generalization to unseen data. Similarly, the YOLOv9e model exhibits a consistent reduction in validation loss, with curves closely following the training patterns, indicating a balanced architecture that effectively resists overfitting. YOLOv9e: a Model performance evaluation metrics, b Training and validation loss curves. YOLOv10 submodels: overview and performance evaluation The performance evaluation of the YOLOv10 submodels was conducted using their respective best.pt weights, which were obtained at the point of highest validation accuracy. The YOLOv10 family consists of six distinct architectures—n, s, m, l, x, and b—each offering varying levels of computational complexity and model capacity. Crucially, this hierarchical structure facilitates flexible deployment across a wide range of platforms, spanning from resource-limited embedded systems to high-performance computing environments. The comparative performance metrics for kiwi fruit detection are presented in Table 6. Table 6Performance metrics of YOLOv10 submodels for Kiwi fruit detection.ModelIterationsPrecisionRecallF1 ScoremAP@0.5mAP@0.5:0.95YOLOv10n1880.8460.8230.8340.9070.587YOLOv10s1330.8650.8510.8580.9220.616YOLOv10m840.8830.8920.8870.9480.637YOLOv10l1050.8900.8830.8860.9450.629 YOLOv10x 84 0.885 0.901 0.893 0.952 0.634 YOLOv10b830.8730.8860.8790.9470.631 Performance metrics of YOLOv10 submodels for Kiwi fruit detection. The YOLOv10 series generally exhibits high detection capability, with each submodel representing a specific point on the accuracy–efficiency trade-off curve. Among the evaluated models, YOLOv10x achieved the highest F1 score (0.893) and the highest mAP@0.5 value (0.952), underscoring its superior precision–recall balance and robust localization accuracy. Notably, the YOLOv10m model also performed competitively (F1 = 0.887, mAP@0.5 = 0.948) while achieving rapid convergence in only 84 iterations, highlighting its excellent training efficiency and well-balanced architecture. The YOLOv10l model delivered comparably strong results (F1 = 0.886), positioning it as a reliable intermediate option between medium and high-capacity models. Furthermore, the lightweight variant YOLOv10s demonstrated a favorable balance between accuracy and efficiency, making it highly suitable for real-time or resource-constrained deployments. In contrast, YOLOv10n, despite requiring a prolonged 188 iterations for convergence, yielded a modest F1 score (0.834), suggesting that its reduced parameter count may limit performance in complex scenes involving occlusions. The final variant, YOLOv10b, produced stable and competitive results (F1 = 0.879), offering a versatile alternative for diverse application requirements. In summary, the models can be categorized based on application: YOLOv10s and YOLOv10b are well-suited for embedded implementations requiring high efficiency, whereas YOLOv10m and YOLOv10x provide optimal performance for accuracy-driven systems that are not constrained by strict hardware limitations. These training dynamics are supported by the iteration-based loss curves and performance trends for YOLOv10m and YOLOv10x, which are illustrated in Figs. 8 and 9, respectively. YOLOv10m: a Model performance evaluation metrics, b Training and validation loss curves. YOLOv10x: a Model performance evaluation metrics, b Training and validation loss curves. Figures 8 and 9 illustrate the evolution of the three primary loss components—box loss, classification loss, and distributional focal loss—during the training and validation phases for the YOLOv10m and YOLOv10x models. Both architectures initiate with high loss values during early epochs and exhibit a consistent downward trend, confirming stable convergence. In YOLOv10m, the close proximity of the training and validation loss curves demonstrates strong generalization, with minimal divergence between the two phases. Slight fluctuations, particularly in the box and DFL losses, are attributed to the inherent complexity of dense foliage and occluded fruit samples. Conversely, the YOLOv10x model, due to its higher capacity, exhibits more localized variations in validation loss. These subtle oscillations may signal sensitivity to overfitting or under-generalization in complex orchard scenes. Nevertheless, the model successfully maintains low overall loss values, evidencing effective learning dynamics. For such high-capacity models, the adoption of careful hyperparameter tuning and early stopping is particularly important to preserve optimal performance. YOLOv11 submodels: overview and performance evaluation The performance evaluation of the YOLOv11 submodels was conducted based on the best.pt weights, which correspond to the highest validation accuracy achieved during the training process. The YOLOv11 family encompasses five variants—n, s, m, l, and x—which are characterized by varying architectural depth and parameter density. This range offers a broad performance spectrum across key metrics, including precision, recall, and detection coverage. The comparative performance metrics of the YOLOv11 submodels for kiwi fruit detection are summarized in Table 7. Table 7Performance metrics of YOLOv11 submodels for Kiwi fruit detection.ModelIterationsPrecisionRecallF1 ScoremAP@0.5mAP@0.5:0.95YOLOv11n850.8620.8970.8790.9420.621YOLOv11s720.8680.9180.8920.9530.637YOLOv11m910.8920.9000.8960.9520.635 YOLOv11l 72 0.882 0.908 0.895 0.957 0.634 YOLOv11x1000.8730.8960.8840.9400.631 Performance metrics of YOLOv11 submodels for Kiwi fruit detection. The YOLOv11 series demonstrates strong potential in both accuracy and detection coverage. The YOLOv11m model achieved the highest F1 score 0.896 and precision 0.892, ranking as the top performer overall. Its ability to reach optimal performance in just 91 iterations highlights its training efficiency. The YOLOv11l variant recorded the highest mAP@0.5 0.957, indicating superior detection coverage. The YOLOv11s model achieved the highest recall 0.918, minimizing missed detections. The YOLOv11n model, despite its lightweight design, delivered competitive results F1 score 0.879, mAP@0.5 0.942, making it well-suited for low-resource environments. Conversely, YOLOv11x achieved balanced performance F1 score 0.884 but required greater computational resources due to its architectural complexity. The YOLOv11 family offers a balanced and robust design in terms of architectural flexibility, detection accuracy, and training efficiency. For lightweight, fast, and low-power applications, YOLOv11s and YOLOv11n are recommended, while YOLOv11m and YOLOv11l are preferable in accuracy-prioritized scenarios. Iteration-based loss curves and performance metrics for YOLOv11m and YOLOv11l are illustrated in Figs. 10 and 11. YOLOv11m: a Model performance evaluation metrics, b Training and validation loss curves. YOLOv11l: a Model performance evaluation metrics, b Training and validation loss curves. Figures 10 and 11 indicate that both YOLOv11m and YOLOv11l exhibited stronglearning dynamics during training. All three primary loss components—box loss, classification loss, and distributional focal loss—started at high values in early iterations and consistently decreased over time. This downward trend was observed in both training and validation sets, suggesting no significant overfitting and strong generalization capabilities. For YOLOv11m, the combination of high F1 score and precision confirms that the model effectively learned both class discrimination and precise localization. Similarly, the YOLOv11l model’s consistent loss reduction and convergence to low values indicate that the model accurately captured bounding box positions and edge alignments. The trajectory in classification loss for both models reflects progressively improved label prediction accuracy. These findings demonstrate that YOLOv11m and YOLOv11l are reliable solutions for high-accuracy and stability-demanding object detection applications. Enhanced training process through hyperparameter optimization In this study, hyperparameter optimization was conducted exclusively on the “s” variants of the YOLO models (YOLOv8s, YOLOv9s, YOLOv10s, and YOLOv11s). The primary motivation for this choice lies in the significantly shorter training times and lower computational requirements of these models, which enable more efficient use of available resources. Moreover, the satisfactory baseline performance results obtained in earlier sections demonstrate that these models are responsive to hyperparameter refinements and can achieve higher accuracy levels through optimization. Based on these considerations, the hyperparameter optimization process for each model was carefully planned, systematically executed, and comparatively evaluated. The hyperparameter settings presented in Table 8 are optimized using the model.tune() function provided by the Ultralytics library, and the training process is conducted in accordance with these configurations. This function employs a hybrid and adaptive optimization strategy that combines Random Search, Bayesian Optimization, and Evolutionary (Genetic) search techniques to manage the hyperparameter exploration process. In this way, the broad and complex hyperparameter space is automatically scanned, the search is guided based on the outcomes of each trial, and parameter combinations yielding progressively higher performance are selected. Table 8Hyperparameter training parameters.ParameterValueDescription/PurposeInput Image Size640 × 640Pixel dimensions of the images processed by the model.Mini-Batch Size8Number of samples fed to the GPU in each training iteration.Maximum Iterations300Maximum number of training steps the model undergoes.Early Stopping Criterion20 IterationsTraining stops if no improvement is observed in the chosen performance metric (typically loss) for 20 iterations.Optimization AlgorithmAdamWOptimization method used to update the model weights.DeviceCUDA-enabled GPU (device = 0)Hardware used for training (for fast computation). Hyperparameter training parameters. To evaluate the performance of each hyperparameter combination, model.tune() utilizes a unified performance metric referred to as fitness. The fitness value represents a single performance score constructed as a weighted average of precision, recall, mAP@0.5, and mAP@0.5:0.95. By summarizing the model’s detection performance into a single numerical value, this approach enables the optimization process to effectively identify the most suitable hyperparameter configuration. In this study, a total of 300 independent trials were executed by the tuner. Each trial operated as a short training cycle designed to evaluate the performance of a specific hyperparameter set, and the resulting fitness values are presented in Figs. 12 and 13 using a scatter plot visualization. Fitness score and iteration count for the YOLOv11s model. Fitness score and iteration count for the YOLOv9s model. Fitness scores which initially exhibit a fluctuating pattern, become more stable as the number of iterations increases and display a generally upward performance trend. This observation supports the conclusion that the hyperparameter search process has a positive influence on model performance. Similar patterns were observed for the YOLOv8s and YOLOv10s models, whose fitness scores also became more stable and showed a general increase as the iterations progressed. To analyze how individual hyperparameters influence model performance, the sampled parameter distributions and their corresponding fitness values are illustrated in Figs. 14. and 15. Each subplot displays the relationship between the sampled values of a specific hyperparameter (e.g., lr0, momentum, scale) and the corresponding fitness scores obtained throughout the tuning process. Sample parameter distribution of the YOLOv11s model. A closer examination of the visualized parameter distributions in Figs. 14 and 15 show insight into the behavior of the sampled hyperparameters. In these graphs, the x-axis represents the values of the respective hyperparameters sampled by the model.tune() function, while the y-axis indicates the overall fitness score produced in each trial. The color of the points reflects the magnitude of the fitness score: lighter colors indicate higher performance, whereas darker colors represent lower performance. These multi-scatter plots offer a clear visual means of assessing the impact of each hyperparameter on model accuracy. Sample parameter distribution of the YOLOv9s model. Within the scope of the hyperparameter optimization process, Table 9 presents a comparative summary of the optimal hyperparameter values obtained for the YOLOv8s, YOLOv9s, YOLOv10s, and YOLOv11s models. This table enables the assessment of the distribution of overall performance (fitness) scores produced across different hyperparameter values. Notably, the dense clustering observed in critical parameters such as lr0, momentum, weight_decay, and scale indicates that specific ranges of these hyperparameters are associated with higher overall performance scores. Table 9Optimal hyperparameter ratios for YOLOv8s, YOLOv9s, YOLOv10s, and YOLOv11s.HyperparameterYOLOv8sYOLOv9sYOLOv10sYOLOv11sDescriptionlr00.009810.005660.001860.00521Initial learning ratelrf0.026270.005460.006410.0205Final learning-rate decay factormomentum0.859400.8450.700000.73652Optimization momentum parameterweight_decay0.000490.000780.000440.00073L2 regularization coefficientwarmup_epochs2.178333.83.710581.2428Learning-rate warm-up durationwarmup_momentum0.727680.4940.85780.34789Initial momentum during warm-upbox6.2018111.19.047917.1372Box regression loss coefficientcls0.205590.280.574820.35725Classification loss coefficientdfl1.105060.7860.733310.41271Distribution Focal Loss coefficienthsv_h0.01180.01220.004550.00667HSV hue jitter amounthsv_s0.90.90.830090.50742HSV saturation jitterhsv_v0.312330.2570.291420.48209HSV value jitterdegrees0.000.00.0Rotation angletranslate0.072240.08850.061470.06738Image translation ratioscale0.161450.2730.793760.19996Random scaling ratioshear0.000.00.0Shear transformation amountperspective0.000.00.0Perspective distortion ratioflipud0.000.00.0Vertical flip probabilityfliplr0.184870.2570.182890.25505Horizontal flip probabilitybgr0.000.00.0BGR channel shift probabilitymosaic1.00.660.824340.73465Mosaic augmentation ratiomixup0.000.00.0MixUp augmentation ratiocopy_paste0.000.00.0Copy-Paste augmentation ratio Optimal hyperparameter ratios for YOLOv8s, YOLOv9s, YOLOv10s, and YOLOv11s. This analysis demonstrates that certain hyperparameter combinations possess the potential to significantly enhance model performance. After the optimization process was completed, the most suitable hyperparameter set for each model was identified and stored in .yaml configuration files. Subsequently, each model was retrained using these optimized hyperparameters. The training process was initiated by loading the final optimized model weights through the Ultralytics YOLO class. The parameters stored in best_hyperparameters.yaml were then imported as a Python dictionary and dynamically passed to the training function. During model training, the AdamW optimization algorithm was employed, with an input image size of 640 × 640 pixels, a mini-batch size of 8, and a total of 250 iterations. The training procedure was carried out on a CUDA-enabled GPU for the single-class object detection task. Through this approach, each model was retrained using its most effective hyperparameter configuration, resulting in more stable, accurate, and consistent outcomes compared to the default settings. Performance analysis after hyperparameter optimization for YOLOv (8s, 9s, 10s, 11s) models The results obtained from the hyperparameter optimization process demonstrate that the “s” variants of the YOLO models deliver high accuracy and balanced performance metrics. Table 10 presents the final performance achieved after retraining each model with its optimized hyperparameter configuration. Table 10Detection metrics for YOLO models retrained with hyperparameter optimization.ModelIterationsPrecisionRecallF1 ScoremAP@0.5mAP@0.5:0.95YOLOv8s1930.9010.8650.8830.9400.621YOLOv9s2450.8760.8950.8860.9440.634YOLOv10s2160.8970.8630.8790.9440.637 YOLOv11s 189 0.883 0.899 0.891 0.956 0.644 Detection metrics for YOLO models retrained with hyperparameter optimization. Retraining the models with the optimal hyperparameters resulted in smoother learning curves and improved performance in key metrics such as precision, recall, F1 score, and mAP. Notably, the YOLOv11s model achieved an F1 score of 0.891 and mAP@0.5 of 0.956, highlighting its outstanding detection accuracy. Its high recall 0.899 minimized missed detections, while its precision 0.883 reduced false positives, enhancing overall detection reliability. The YOLOv9s model exhibited balanced and consistent performance, achieving an F1 score of 0.886 and mAP@0.5 of 0.944. Its strong recall confirmed that it successfully captured the majority of detectable objects. Similarly, the YOLOv10s model achieved an mAP@0.5 of 0.944 and an F1 score of 0.879, demonstrating consistent detection capability. The training consistency of this model suggests that certain hyperparameter sets positively influence its learning dynamics. The YOLOv8s model achieved the highest precision 0.901 and an F1 score of 0.883, effectively minimizing false positives. Although its training extended to 193 iterations, the final accuracy levels justify the prolonged training process. In summary, the hyperparameter optimization allowed each model to be retrained with settings tailored to its architecture, resulting in meaningful improvements in overall performance. Among all, the YOLOv11s model demonstrated superior results across all metrics, representing the most successful outcome of the hyperparameter optimization process. Iteration-based loss curves and performance trends for the YOLOv11s model are shown in Fig. 16. YOLOv11s Hyperparameter optimization: (a) Model performance evaluation metrics, (b) Training and validation loss curves. Figure 16 illustrates the changes in box loss, classification loss, and distributional focal loss for the YOLOv11s model during both training and validation phases. In the early iterations, all loss components start at relatively high values but steadily decline over time, indicating consistent learning dynamics. The consistent decrease in box loss reflects the model’s improving ability to accurately localize object boundaries. Although occasional fluctuations were observed in classification loss during validation, the overall performance remained stable, indicating that temporary challenges in class prediction did not degrade the final outcome. The low behavior of distributional focal loss in both training and validation phases suggests that the model is able to optimize bounding box coordinates with high precision. The close alignment of training and validation losses indicates a low risk of overfitting and confirms that the model maintained strong generalization capability. This shows that hyperparameter optimization contributed not only to accuracy metrics but also to training stability and model balance. The balanced convergence observed in the YOLOv11s model supports its reliability for complex visual tasks such as agricultural object detection. Figure 17 illustrates the detection results produced by the optimized YOLOv11s model, demonstrating its ability to accurately identify and localize kiwi fruits under complex orchard conditions. As shown in Fig. 17, the model successfully detects a large number of densely clustered fruits despite significant variations in illumination, color intensity, and background clutter. The bounding box predictions remain stable even in regions with heavy leaf occlusion, overlapping fruits, and shadowed areas—scenarios that typically challenge real-time fruit detection systems. These results highlight the robustness gained through the hyperparameter optimization process, enabling the model to maintain high detection consistency across diverse environmental conditions. Overall, the visual outputs confirm that the optimized YOLOv11s model is capable of delivering reliable and precise fruit localization in real-world agricultural settings. Detection outputs of the YOLOv11s model after hyperparameter optimization. Table 11 shows an integrated comparison of the YOLOv8, YOLOv9, YOLOv10, and YOLOv11 model families alongside their HPO-optimized variants. The results reveal clear performance patterns across model scales and generations. As expected, larger configurations (m, l, x) consistently achieve higher F1 scores and mAP values, reflecting their enhanced representational capacity. Within the YOLOv8 series, the YOLOv8x model provides the strongest overall accuracy, achieving the highest mAP@0.5 (0.956) and a competitive F1 score (0.891). In the YOLOv9 family, YOLOv9e stands out with the highest mAP@0.5:0.95 (0.645), demonstrating the benefit of the PGI-based optimization strategy. Table 11Comparison of YOLOv8–YOLOv11 models with HPO.ModelIterationsPrecisionRecallF1 ScoremAP@0.5mAP@0.5:0.95YOLOv8n920.8850.8910.8880.9510.620YOLOv8s920.8750.9040.8890.9530.627YOLOv8m560.8790.8730.8760.9420.628YOLOv8l610.8740.8750.8750.9460.636YOLOv8x560.8830.8990.8910.9560.640YOLOv9t810.8760.8850.8800.9480.623YOLOv9s680.8640.8950.8790.9460.630YOLOv9m620.9000.8700.8850.9500.635YOLOv9c840.8800.8780.8790.9410.633YOLOv9e660.8870.8900.8890.9510.645YOLOv10n1880.8460.8230.8340.9070.587YOLOv10s1330.8650.8510.8580.9220.616YOLOv10m840.8830.8920.8870.9480.637YOLOv10l1050.8900.8830.8860.9450.629YOLOv10x840.8850.9010.8930.9520.634YOLOv10b830.8730.8860.8790.9470.631YOLOv11n850.8620.8970.8790.9420.621YOLOv11s720.8680.9180.8920.9530.637YOLOv11m910.8920.9000.8960.9520.635YOLOv11l720.8820.9080.8950.9570.634YOLOv11x1000.8730.8960.8840.9400.631YOLOv8s-HPO1930.9010.8650.8830.9400.621YOLOv9s-HPO2450.8760.8950.8860.9440.634YOLOv10s-HPO2160.8970.8630.8790.9440.637YOLOv11s-HPO1890.8830.8990.8910.9560.644 Comparison of YOLOv8–YOLOv11 models with HPO. The YOLOv10 models show robust performance, particularly YOLOv10m, YOLOv10x, and YOLOv10b, which benefit from the Fast–Slow architectural design and NMS-free detection pipeline. The latest generation, YOLOv11, exhibits consistently strong results across all scales, with YOLOv11s, YOLOv11m, and YOLOv11l achieving some of the highest recall and F1 scores in the entire table, confirming the effectiveness of the hybrid CNN–Transformer feature extraction mechanism. The HPO-enhanced variants further show that systematic hyperparameter tuning provides measurable performance improvements, especially in the YOLOv11s-HPO model, which reaches one of the highest mAP@0.5:0.95 values (0.644). Similar gains, though smaller in magnitude, are observed in YOLOv9s-HPO and YOLOv10s-HPO. Overall, the results indicate that newer YOLO versions deliver more stable and accurate detection performance, while larger model scales and targeted HPO strategies offer additional improvements. These findings justify the inclusion of multiple YOLO families in the study and provide a clear understanding of their comparative strengths in object detection tasks. Kiwi detection on embedded systems using YOLO architecture In the experimental implementation, the YOLO models are demonstrated directly on the Jetson TX2 platform to evaluate their low-latency inference performance in embedded environments. The tests are conducted using a static test dataset comprising 40 images and 1,936 bounding box annotations, with key metrics such as per-image inference time analyzed. A captured frame demonstrates the YOLO model performing kiwi detection on the Jetson TX2, where detection results are overlaid on the image. The terminal output displays the per-image inference time values as the model processes the input data. The physical setup of the test system, showing the Jetson TX2 operating with a connected display, keyboard, and peripheral connections, is also illustrated. This experimental setup highlights that, for YOLO deployment on embedded systems, evaluation is determined not only by accuracy but also by inference time, hardware compatibility, and the potential for low-latency processing. Figure 18 presents visual examples of the embedded test environment used for these evaluations. Embedded test environment for real-time kiwi detection on Jetson TX2. Inference time performance of YOLO models on Jetson TX2 The inference time performance of all YOLO model families (YOLOv8–YOLOv11) is evaluated on the NVIDIA Jetson TX2 platform to determine their suitability for low-latency embedded applications. Table 12 summarizes the average, minimum, and maximum inference times for each model variant. Table 12Inference time performance of YOLO models on Jetson TX2.ModelAvg. Time (s)Min (s)Max (s)YOLOv8n1.3581.3102.533YOLOv8s3.9243.8466.104YOLOv8m9.8299.69613.172YOLOv8l19.79719.55123.818YOLOv8x27.90427.46634.448YOLOv9t1.5471.4933.131YOLOv9s4.1234.0245.952YOLOv9m10.34110.21512.964YOLOv9c13.67713.52816.672YOLOv9e26.07025.59931.657YOLOv10m1.5121.4643.006YOLOv10s3.8003.7135.797YOLOv10l8.8638.71911.496YOLOv10b13.33713.17816.234YOLOv10x21.32020.87225.599YOLOv11n1.2481.1942.733YOLOv11s3.3153.2305.365YOLOv11m9.5729.38512.610YOLOv11l12.31912.13815.721YOLOv11x26.19024.72732.732 Inference time performance of YOLO models on Jetson TX2. Across all YOLO families, lightweight variants (n, s, t, m) demonstrate the best balance between inference time and resource constraints, making them suitable for near real-time agricultural detection on Jetson TX2. The fastest overall models are YOLOv11n (1.248 s), YOLOv10m (1.512 s), and YOLOv9t (1.547 s), all meeting the requirements for low-latency operation. Mid-tier models (YOLOv8m, YOLOv9m, YOLOv10l, YOLOv11m) exceeded the 8–10 s inference time threshold, which pushes the limits of low-latency feasibility but offers higher accuracy for moderate-latency applications. Larger models (YOLOv8l/x, YOLOv9e, YOLOv10x, YOLOv11x) consistently exceeded 20 s per inference, rendering them impractical for low-latency use on Jetson TX2 without substantial optimization. Kiwi detection on embedded systems: performance evaluation of Hyperparameter-Optimized YOLOv (8s, 9s, 10s, 11s) models In this section, the impact of hyperparameter optimization on the latest lightweight YOLO models—YOLOv8s, YOLOv9s, YOLOv10s, and YOLOv11s—is evaluated on the NVIDIA Jetson TX2 platform. Focusing on low-latency inference metrics, the comparative analysis highlights the performance variations observed after optimization. The results are summarized in Table 13. Table 13Inference time of hyperparameter-optimized YOLOv(8–11 s) models on Jetson TX2.ModelAvg. Time (s)Min (s)Max (s)YOLOv8s4.1004.0225.882YOLOv9s4.1944.0966.013YOLOv10s3.8433.7505.850 YOLOv11s 3.333 3.248 5.417 Inference time of hyperparameter-optimized YOLOv(8–11 s) models on Jetson TX2. Following hyperparameter optimization, the YOLOv11s model achieved the shortest average inference time of 3.333 s, establishing itself as the fastest among all tested models. This result demonstrates that YOLOv11s benefited most from optimization, delivering an effective balance of speed and accuracy for low-latency embedded applications. The YOLOv10s model, with an average inference time of 3.843 s, also performed competitively, making it a strong alternative for near real-time agricultural image processing systems when optimized. YOLOv8s achieved a 4.100-second average inference time, maintaining acceptable performance; however, despite the gains from optimization, it remains secondary compared to the higher-performing YOLOv10s and YOLOv11s models. The YOLOv9s model recorded the longest inference time at 4.194 s, yet it demonstrated competitive accuracy, making it suitable for scenarios where precision is prioritized over strict latency constraints. After hyperparameter optimization, YOLOv11s consistently emerged as the best-performing model in terms of both inference speed and accuracy, making it the most suitable configuration for low-latency kiwi detection on resource-limited embedded platforms such as the Jetson TX2. Figure 19 illustrates the detection output of the hyperparameter-optimized YOLOv11s model running on Jetson TX2. Detection output of the hyperparameter-optimized YOLOv11s model on Jetson TX2. The optimized YOLOv11s model demonstrates a well-balanced structure for low-latency agricultural detection tasks in embedded systems, while the other optimized models provide flexibility for specific application scenarios where either latency or accuracy requirements may differ. On embedded platforms such as the Jetson TX2, inference times below 5 s are frequently classified in the literature as “near real-time”59. In agricultural image processing applications, camera systems often operate at low frame rates (approximately 1 FPS or lower); therefore, an inference time of 3–4 s is generally considered sufficient for most use cases60. For instance, in a cotton detection study implemented on the Jetson TX2, an inference time of 2.5 s was reported, and the system was evaluated as near real-time61. Similarly, other works have demonstrated Jetson TX2-based systems achieving 13–14 FPS, confirming their suitability for time-critical agricultural applications61. These findings indicate that the inference times obtained in this study—ranging from 1.2 to 4.1 s—are consistent with the literature and confirm the practical applicability potential of the proposed models in embedded agricultural detection scenarios. The comparative results presented in Table 14 provide a broader context for evaluating the performance of the proposed YOLOv11s model relative to existing approaches in the current literature. The comparative analysis of recent studies highlights the rapid evolution of lightweight and attention-enhanced YOLO architectures for horticultural detection tasks. Models such as GhostNet-YOLOv4 and S-YOLOv4-tiny demonstrate that computational efficiency can be achieved without substantial loss of accuracy, while more advanced designs—such as EDT-YOLOv8n, YOLOv8-BFW, and YOLOv10-Kiwi—introduce architectural refinements that improve robustness in complex orchard environments. Attention-based variants (e.g., CBAM-YOLOv8 and YOLOv8 + SE) strengthen performance in crowded or small-object scenarios. Notably, segmentation-based methods like Mask R-CNN continue to achieve exceptionally high accuracy, albeit with greater computational cost. Compared with these works, the present study provides a unique contribution by systematically benchmarking YOLOv8 through YOLOv11 across multiple scales, integrating hyperparameter optimization, and evaluating embedded performance on the Jetson TX2. While some previous models achieve higher accuracy or faster frame rates individually, none offers a holistic comparison across four YOLO generations combined with an embedded-system analysis. The optimized YOLOv11s model presented here therefore fills an important gap by demonstrating a practical and well-balanced solution for near-real-time kiwi detection under real orchard conditions. Table 14Comparative summary of related Studies.StudyModelEnvironment / DatasetPerformance (mAP / Accuracy)Gao et al.13GhostNet-YOLOv4Natural orchard0.9307Yang et al.5S-YOLOv4-tinyOrchard environment0.8975Chen et al.8EDT-YOLOv8nComplex orchard environment~ 0.94Ulutaş49Mask R-CNN (ResNet50)Orchards in Samsun, Turkey0.9848Shan et al.62YOLOv8-BFWKiwi dataset0.969 mAPRen et al.63YOLOv10-KiwiPergola-trained orchard0.936 mAP@50, 74 FPSXia et al.64CBAM-YOLOv8 + Focal-EIoU1224 kiwi images0.942 mAP@50Lee et al.65YOLOv11m (Optimized C47)22k tomato disease dataset0.99197 mAP@0.5Ajayi et al.66YOLOv8UAV-based multi-crop classification96–98% AccuracySolimani et al.67YOLOv8 + AttentionTomato flower–fruit–node detection94–98% mAPRamos et al.16Optimized YOLOv8Smoke / wildlife monitoringmAP@0.5:0.95 improved by + 5.09% Our study YOLOv11s (Hyperparameter-Optimized) Orchards in Samsun, Turkey 0.956 mAP@0.5 (3.33 s inference on Jetson TX2) Comparative summary of related Studies. Discussion The findings of this study highlight the trade-offs between architectural complexity, detection accuracy, generalization capability, and computational efficiency when deploying YOLO-based models for fruit detection in real agricultural environments. A consistent pattern emerged across all YOLO families (v8–v11): while larger models (e.g., YOLOv8x, YOLOv9e, YOLOv10x, YOLOv11x) achieved slightly higher accuracy on the desktop training system, their inference times—often exceeding 20 s per image—render them unsuitable for deployment on constrained embedded platforms such as the Jetson TX2. In contrast, lightweight variants (n, s, and m) delivered a more favorable balance. Among these, the hyperparameter-optimized YOLOv11s model achieved the best overall performance with an mAP@0.5 of 0.956 and an average inference time of 3.33 s, aligning with definitions of near-real-time performance in agricultural monitoring applications. The hyperparameter optimization procedure played a significant role in improving the stability and robustness of the lightweight models. The optimization resulted in smoother convergence curves, reduced overfitting tendencies, and measurable gains in F1 score and mAP across all optimized models. This demonstrates that substantial performance improvements can be obtained through systematic tuning, even without architectural modification—an important insight for embedded AI deployments where model complexity must remain minimal. The YOLOv11s model benefited most from hyperparameter tuning, showing improvements across precision, recall, and mAP, confirming that lightweight architectures respond more sensitively to optimization strategies due to their narrower parameter space. In addition to accuracy, inference latency remains a fundamental determinant of model suitability for field deployment. While robotics-driven agricultural systems generally require frame rates above 5 FPS, several studies in precision agriculture have classified inference times between 1 and 4 s as acceptable for static frame analysis, low-frequency monitoring, or stop-and-go operation. The inference times obtained in this study fall within this range, suggesting that optimized lightweight YOLO models can support practical near-real-time fruit detection tasks such as yield estimation, decision support, and orchard monitoring—even though they do not yet meet strict robotic real-time thresholds. Future hardware acceleration techniques such as TensorRT-based INT8 quantization or structured pruning are likely to reduce latency further, potentially enabling true real-time operation on the Jetson TX2. Despite the promising results, several limitations must be acknowledged. First, the dataset—although collected under natural orchard conditions with significant variability in lighting, occlusion, and canopy density—remains relatively small. This may limit generalization to orchards with different training systems, varieties, or environmental conditions. Expanding the dataset across multiple seasons, weather scenarios, and orchard structures would strengthen robustness. Second, the real-time embedded evaluation was performed on pre-collected test images rather than continuous video input. Although still valid for inference benchmarking, future work should test the models under online streaming conditions to more accurately assess end-to-end system latency. Third, while this study focused on YOLO architectures due to their popularity and efficiency, alternative models such as DETR variants, EfficientDet, MobileNet-SSD, or foundation models like Grounding DINO and SAM could provide valuable baselines for more comprehensive benchmarking. Another important observation concerns the computational trade-offs across YOLO versions. YOLOv11 consistently offered improved accuracy and efficiency over earlier architectures, confirming the advantages of its enhanced feature extraction blocks and optimization strategies. However, the performance differences between YOLOv10s, YOLOv11s, and YOLOv9s after hyperparameter optimization were relatively narrow, suggesting that architectural innovation alone is not always the primary determinant of embedded performance. Instead, a careful balance of architecture selection, dataset design, augmentation strategy, and hyperparameter tuning appears to have a greater cumulative effect. Finally, although this study focused on kiwi fruit detection, the methodology and optimization pipeline are highly transferrable to other horticultural crops with similar detection challenges—such as citrus, apples, plums, and peppers. Many agricultural fruits share structural similarities in terms of occlusion patterns, canopy density, and color blending with the surrounding foliage. Therefore, with appropriate retraining and dataset expansion, the optimized YOLO-based framework presented here has strong potential to support a wide range of orchard automation tasks including yield forecasting, selective harvesting, growth monitoring, and robotic perception. Conclusions This study presented a comprehensive evaluation of YOLOv8–YOLOv11 architectures for kiwi fruit detection under natural orchard conditions and on an embedded platform. By training all model variants under a uniform experimental setup and applying hyperparameter optimization to the lightweight “s” models, the research systematically examined the trade-offs between detection accuracy, model complexity, and embedded inference performance. The results demonstrated that while larger models achieved slightly higher accuracy on high-performance hardware, their latency on embedded devices makes them impractical for field deployment. In contrast, the optimized YOLOv11s model achieved the best overall balance, reaching an mAP@0.5 of 0.956 with an average inference time of 3.33 s on the Jetson TX2, establishing it as the most suitable near-real-time solution among all evaluated models. The findings emphasize the importance of targeted hyperparameter optimization, which significantly improved the performance and stability of the lightweight architectures without increasing computational cost. These results demonstrate that substantial gains can be achieved not only through architectural innovation but also through carefully designed training strategies—an especially valuable insight for embedded agricultural AI applications. The methodology and evaluation pipeline used in this study provide a transferable framework that can be applied to similar fruit detection problems across other horticultural domains. Despite the strong results, several limitations must be acknowledged. The dataset, although collected under realistic orchard conditions, remains limited in size and environmental diversity. Real-time inference tests were conducted on pre-collected images rather than continuous video streams, and the analysis focused primarily on YOLO-based architectures without benchmarking against alternative lightweight detectors. Addressing these limitations will strengthen system generalization and practical applicability. Future research will explore the integration of advanced optimization techniques, including TensorRT-based quantization, pruning, and knowledge distillation, to reduce latency and enable true real-time performance on edge devices. Expanding the dataset across different orchards, seasons, and environmental conditions, as well as incorporating multi-modal sensing (e.g., depth, thermal imaging), will also enhance robustness. Additionally, evaluating non-YOLO models and transformer-based architectures may provide new insights into the performance boundaries of embedded agricultural vision systems. This study demonstrates that optimized lightweight YOLO architectures—particularly YOLOv11s—offer a viable and efficient solution for near-real-time kiwi detection in resource-constrained environments, supporting the broader adoption of AI-driven automation and intelligent decision-making in precision agriculture. Author contributions B.K. prepared and pre-processed the dataset, implemented and trained the YOLO models, and performed the experimental analyses. M.E.S. designed the study, supervised the research, conducted hyperparameter optimization, and wrote the main manuscript text. Both authors contributed to the discussion of results, prepared figures and tables, and reviewed the manuscript. Data availability Data is available on public as given in [https://github.com/essgungor/kivi](https:/github.com/essgungor/kivi) . Declarations Competing interests The authors declare no competing interests. Ethics approval Not applicable. 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 Data is available on public as given in [https://github.com/essgungor/kivi](https:/github.com/essgungor/kivi) .