In recent years, artificial intelligence and deep learning technologies have been gradually applied to various agricultural domains, such as crop cultivation, harvesting, and disease detection (Kamilaris and Prenafeta-Boldú, 2018). Farmers have begun using smartphones to detect crop diseases and pests (Liu et al., 2020). However, this field still faces numerous challenges and development opportunities.

Traditional agricultural disease detection and identification methods primarily rely on manual feature extraction. While these methods have achieved certain research results (Camargo and Smith, 2009; Shekhawat and Sinha, 2020), they have apparent limitations. These conventional approaches necessitate substantial professional expertise and a deep reservoir of knowledge, are inherently subjective, and often neglect valuable attributes that are challenging to identify with unaided human perception. Furthermore, when confronted with voluminous data in authentic natural settings, the precision of traditional methods frequently deteriorates significantly (Buja et al., 2021; Wiesner-Hanks et al., 2019).

In contrast, deep learning technologies, with their potent feature representation capacities, can autonomously extract features from voluminous multi-type disease data, thereby substantially enhancing detection performance (Saleem et al., 2021; Bhattacharya et al., 2022). AI has made disease detection more automated, efficient, precise, and reliable, primarily manifested in the following aspects:

The efficacy of deep learning models in plant disease detection is intrinsically tied to the quality of the training data. Currently, the creation of datasets for agricultural disease detection models can be broadly classified into two primary categories: natural environment photography (with background) and ideal environment (without background), as depicted in Figure 1 . Images captured in natural environments present complex backgrounds, which can enhance the robustness and generalization capabilities of trained models. Conversely, images acquired in ideal environments lack background distractions, but the resulting models often struggle to achieve satisfactory detection performance in real-world scenarios.

Given the intricate relationship between agricultural diseases and factors like cultivation practices, management strategies, and climate fluctuations, prevailing open-source tomato disease datasets largely depend on laboratory samples, exemplified by AI Challenger 2018, Kaggle, and PlantVillage. Recognizing the considerable time and effort necessary to accumulate a substantial quantity of natural environment samples, numerous agricultural disease detection model studies predominantly leverage open-source ideal environment samples for training purposes. Although models developed based on ideal environment samples have achieved notable outcomes in laboratory settings, many have not undergone validation in natural environments. Existing research suggests that these models are primarily well-suited for scenarios where disease-affected areas constitute a significant portion of the image but may encounter difficulties in handling complex backgrounds, lighting variations, changes in shooting angles, and diverse lesion sizes within natural scenes.

For instance, Bora et al. (2023) developed a system capable of detecting diseases on tomato leaves, stems, fruits, and roots with remarkable accuracy rates of 99.84%, 95.2%, 96.8%, and 93.6%, respectively. Zhang et al. (2023) introduced the M-AORANet model, which demonstrated exceptional recognition accuracy of 96.47% on a dataset comprising 3,123 tomato leaf images. Sunil et al. (2023) employed a Multi-level Feature Fusion Network (MFFN) to achieve an impressive external test accuracy of 99.83% on publicly available tomato disease datasets. Although these models exhibited exceptional performance in controlled environments, their primary limitation lies in their inability to pinpoint the exact location of lesions within images, hindering their direct application in real-world agricultural settings.

Although models trained on natural environment data more accurately reflect real-world conditions, the majority of research continues to concentrate on model development, refinement, and structural analysis using personal computers. This neglects the critical need for lightweight and highly accurate solutions in practical applications.

Several prior studies have explored the use of deep learning for plant disease detection in various agricultural settings. Li et al. (2019) developed a mobile application for early detection of tomato late blight, demonstrating the potential for smartphone-based disease diagnosis. Sun et al. (2021) proposed the MEAN-SSD model, achieving an accuracy of 83.12% for apple leaf disease detection while maintaining real-time processing speeds (12.53 frames per second). Zhang K. et al. (2021) incorporated skip connections within the Faster R-CNN architecture, achieving an accuracy of 83.34% on a custom dataset of soybean disease images. Similarly, Chen et al. (2021) constructed a model for cucumber leaf disease detection with an accuracy of 85.52%. Finally, Dananjayan et al. (2022) demonstrated the effectiveness of YOLOv4 for rapid and accurate detection of citrus leaf diseases. Qi et al. (2022) introduced an enhanced SE-YOLOv5s network model that achieved an impressive 91.07% accuracy on the tomato disease test set. While these studies demonstrated real-time disease recognition capabilities, the models developed for individual agricultural diseases face challenges in widespread deployment due to the fluctuating nature of disease occurrences.

Machine vision detection of tomato diseases faces significant obstacles in actual planting environments, including complex growing conditions, multiple disease types, and subtle symptom variations (Barbedo, 2018; Kadry, 2021; Fuentes et al., 2021), which impose exceptionally high demands on the multi-feature and cross-scale extraction capabilities of detection algorithms. While the YOLO series models have garnered widespread adoption for their swift and precise detection capabilities, there remains potential for enhancement in feature extraction and detection accuracy within complex environments.

Considering the current research status and challenges, this study proposes a tomato disease detection method based on a refined YOLOv10 architecture. By meticulously analyzing tomato disease types and image characteristics, the algorithm is iteratively enhanced and experimentally validated, aiming to fulfill the accuracy and speed requirements of intelligent tomato disease detection and reduce manual diagnosis costs. The innovations of this study are primarily as follows:

In this study, we utilized a custom-built tomato disease dataset aimed at capturing disease features in real agricultural environments, reflecting the imbalanced nature of disease occurrence in actual scenarios. The data collection process was rigorously controlled to ensure quality and representativeness. Table 1 summarizes the key parameters of data collection.

The data acquisition process prioritized capturing a broad spectrum of environmental conditions, encompassing diverse lighting scenarios, angles of view, and background elements, such as leaves, weeds, and soil. This diversity is crucial for improving the model’s generalization ability and applicability in real-world conditions. Each image was accompanied by rich metadata, including environmental temperature, precise location, and timestamp, providing valuable context for subsequent analysis and model training. Figure 2 shows representative samples of the various tomato diseases in the dataset, visually illustrating the characteristics and complexity of different diseases.

Our dataset reflects the natural distribution of disease occurrence in actual agricultural environments, thus exhibiting significant class imbalance. Table 2 details the sample count for each category and its proportion in the overall dataset.

This imbalanced distribution reflects the relative frequency of various diseases in actual agricultural production, providing us with a realistic challenge scenario. In particular, the sample sizes for leaf mold and gray mold are notably smaller than other categories, highlighting the importance and difficulty of identifying rare diseases in practical applications.

To address the dataset’s class imbalance, we employed a stratified random sampling technique. This method ensured that the distribution of each category in the training, validation, and testing sets accurately mirrored the original dataset. A detailed breakdown of the dataset division is provided in Table 3 .

This division method ensures that each subset contains balanced representations of all categories while preserving the imbalanced characteristics of the original dataset, contributing to stable performance and reliable evaluation of the model in real-world scenarios.

Accurate data annotation is indispensable for guaranteeing the efficacy of model training. This research employed the widely utilized open-source annotation tool Labellmg to prepare data for object detection purposes. The annotation process is visually depicted in Figure 3 .

Given the dataset’s uneven distribution, we implemented focused data augmentation techniques designed to address class imbalance concerns and bolster the model’s capacity to identify uncommon classes.

Although the C2f module in YOLOv10 enhances feature extraction capabilities through the Bottleneck structure, its equal treatment of all channels and positional information introduces a significant amount of irrelevant interference. This results in suboptimal performance when handling multi-scale small targets and complex backgrounds in tomato disease images. Additionally, in the YOLOv10 model, the backbone network extracts deep features through multiple down-sampling convolution layers.

Although this multi-stage downsampling enhances the model’s capacity to handle large targets and intricate backgrounds, it also leads to a significant reduction in small target features. To mitigate the loss of these crucial details during detection, the neck network utilizes multiple upsampling operations to restore feature map resolution and integrate features from various levels, thereby improving the model’s ability to detect targets of varying sizes. However, this alternating process of downsampling and upsampling also results in excessive layer stacking within the backbone and neck networks, increasing the model’s parameter count and computational complexity, making it challenging to meet real-time detection requirements in practical applications. To overcome these aforementioned challenges, given the nature of multi-scale target detection of tomato diseases in intricate environments, this study proposes a streamlined target detection algorithm, TomatoGuard-YOLO, as illustrated in Figure 5 .

This model is an enhancement of version n of YOLOv10 (referred to as YOLOv10 unless otherwise specified), featuring lightweight designs for both the backbone and neck networks. Initially, a novel feature extraction and fusion module, termed the Multi-Path Inverted Residual Unit (MPIRU), is devised. Subsequently, the proposed Dynamic Focusing Attention Framework (DFAF) is integrated into the C2f module, resulting in the C2f-DFAF module. Subsequently, all C2f modules within the foundational YOLOv10 backbone and neck networks are supplanted with MCIR and C2f-DFAF. Furthermore, the backbone network’s downsampling operations have been curtailed to retain a greater quantity of feature information. Concurrently, the upsampling and feature concatenation procedures within the neck network have been streamlined to diminish the number of layers and complexity, thereby further reducing computational expenses. In conclusion, the loss function has been meticulously refined to Focal-EIoU, effectively rectifying the deficiencies of the original loss function and bolstering the model’s capacity to concentrate on a variety of samples.

The subsequent sections elucidate each improved module, aiming to elevate the model’s performance in detecting tomato diseases in intricate agricultural environments, particularly in handling small targets, complex backgrounds, and class imbalance challenges.

To ensure the reproducibility and reliability of the experimental results, this study meticulously records the experimental environment and key parameter settings. Tables 6 , 7 list the hardware and software configurations, as well as the core parameters for model training.

This experiment employs cutting-edge hardware and software configurations, particularly the NVIDIA A100 GPU, whose powerful computing capabilities significantly enhance the efficiency of large-scale model training. The deep learning framework used is PyTorch, a popular open-source platform that seamlessly integrates with the hardware to optimize performance. Additionally, the selected versions of CUDA and cuDNN ensure full utilization of GPU acceleration and provide robust support for neural networks.

These parameters were chosen based on multiple experiments and optimizations. A batch size of 64 effectively utilizes the parallel capabilities of multi-GPU setups. The AdamW optimizer, combined with the cosine annealing strategy, enhances model convergence and mitigates potential overfitting issues during training. To ensure thorough training, an early stopping mechanism is introduced to prevent overfitting on the validation set.

This study employs a set of indicators, such as mAP, Model Size (in MB), Number of Parameters (in MB), and Frames Per Second (FPS). These indicators not only reflect the model’s detection accuracy but also provide insights into its computational complexity and real-time performance. The formulas for these indicators are as follows:

Where TP represents True Positives, FP represents False Positives, and FN represents False Negatives. Precision measures the proportion of true positive samples among all samples predicted as positive, reflecting the accuracy of the model’s positive predictions. Recall, on the other hand, indicates the proportion of true positive samples correctly identified by the model among all actual positive samples, assessing the model’s recognition ability. The mean Average Precision (mAP) is calculated by averaging the Average Precision (AP) for each class, with K denoting the total number of classes. It serves as a comprehensive performance metric in object detection tasks, demonstrating the model’s effectiveness across multiple categories. The F1-score, which is the harmonic mean of Precision and Recall, offers a balanced evaluation of these metrics, particularly useful when there’s an imbalance between them.

In addition to these accuracy evaluation metrics, this study incorporates the following indicators to assess model efficiency:

Through a comprehensive evaluation of these indicators, this study not only confirms the accuracy and generalization capabilities of the TomatoGuard-YOLO model in tomato disease detection but also examines its computational efficiency and resource consumption, providing robust support for its practical application.

The learning rate is a crucial hyperparameter in training deep learning models, significantly influencing convergence speed, final performance, and stability. To identify the optimal learning rate for the TomatoGuard-YOLO model, we designed multiple comparative experiments, testing initial rates of 0.1, 0.05, 0.01, and 0.001 while keeping other hyperparameters constant. Figure 9 illustrates the trends in the loss function and accuracy for these different learning rates.

As illustrated in the loss function curves in Figure 9A , it is clear that with learning rates of 0.1 and 0.05, the loss function does not exhibit substantial reduction, instead displaying a relatively stable or even fluctuating trend. This suggests that the learning rate is excessively high, resulting in overly large gradient update steps that hinder the model’s convergence to an optimal solution, thereby affecting training outcomes. Conversely, with learning rates of 0.01 and 0.001, the loss function demonstrates a stable and marked downward trend, indicating that the model gradually approaches the optimal solution. Among these, a learning rate of 0.001 yields the most consistent decrease in the loss function, ultimately reaching its lowest point and demonstrating the best convergence.

Figure 9B presents the accuracy changes of the model under different learning rates. It is evident that at learning rates of 0.1 and 0.05, the accuracy curves exhibit significant fluctuations and fail to stabilize at a high level. This aligns with the earlier observation of the loss function’s instability, further confirming that a high learning rate induces training volatility, preventing the model from achieving optimal performance. In contrast, at a learning rate of 0.01, the accuracy rises rapidly, but in the mid-to-late training stages, the accuracy curve begins to display slight fluctuations and tendencies towards overfitting. When employing a learning rate of 0.001, accuracy improves steadily, ultimately reaching the highest value without significant overfitting, resulting in a stable training process with excellent convergence.

Considering the convergence behavior of the loss function, the rate of accuracy improvement, and the model’s stability during training, we ultimately selected 0.001 as the initial learning rate for the TomatoGuard-YOLO model. This learning rate ensures stable convergence while providing sufficient capacity for the model to fully adapt to the complex patterns and variations in the training data, laying a solid foundation for subsequent optimizations.

Based on the experimental environment configuration and parameter settings described earlier, we systematically trained the TomatoGuard-YOLO model for a total of 300 epochs. To comprehensively understand the performance during training, we closely monitored the changes in several key performance indicators, such as the loss function, mAP50, and mAP50:95. Figure 10 shows the trends of these indicators during the training process.

As illustrated in Figure 10 , both the training loss and validation loss exhibit a consistent and marked decline, with the gap between them gradually narrowing. This suggests that the model is effectively adapting to the training data while simultaneously enhancing its generalization capability. Specifically, the box_loss decreases from an initial value of 0.086 to around 0.030, obj_loss drops from 0.038 to approximately 0.027, and cls_loss significantly reduces from an initial 0.071 to 0.007. These results indicate notable enhancements in the model’s localization accuracy (box_loss), detection confidence (obj_loss), and classification performance (cls_loss).

In terms of detection accuracy, the mAP50 and mAP50:95 indicators show a rapid increase in the early training stages, followed by a gradual slowdown and eventual stabilization in the later phases. Ultimately, mAP50 reaches 94.23%, while mAP50:95 stabilizes at 72.52%, indicating that the TomatoGuard-YOLO model possesses excellent object detection capabilities across different IoU thresholds.

Additionally, during the training process, we observed that the model’s ability to recognize various types of tomato diseases gradually improved. Notably, it maintained high detection accuracy even for disease types with relatively fewer samples. Through gradual adjustments and optimizations, the TomatoGuard-YOLO model achieved outstanding performance in tomato disease detection.

To systematically evaluate the effectiveness of our proposed improvements, we conducted comprehensive ablation experiments on the core enhancement modules based on the YOLOv10 model. Through eight carefully designed comparative experiments with different module combinations, we assessed the impact of the MPIRU module, c2f-DFAF module, and Focal-EIoU loss function on tomato disease detection performance across multiple dimensions, including detection accuracy, model size, and computational overhead. Table 8 presents detailed key performance indicators under various module combinations.

The experimental results reveal the significant impact of each enhancement module on model performance. The standalone introduction of the MPIRU module led to substantial improvements in detection accuracy, with mAP50 increasing from 84.15% to 91.82% and mAP50:95 rising from 61.22% to 68.71%. More importantly, this performance enhancement was accompanied by a significant reduction in model complexity, with model size decreasing from 5.61MB to 2.62MB and parameter count reducing from 2.73MB to 0.93MB. These results convincingly demonstrate the MPIRU module’s effectiveness in simultaneously improving detection accuracy and achieving model lightweighting.

The integration of the c2f-DFAF module, through optimization of the C2f structure, further enhanced model performance, achieving an mAP50 of 88.92% and mAP50:95 of 65.81%. While the parameter count increased slightly (from 2.73MB to 2.74MB), the significant performance improvements fully justify this optimization. The module demonstrated excellence in enhancing feature hierarchy capture and multi-scale feature fusion, effectively improving the model’s feature expression capabilities.

The application of the Focal-EIoU loss function exhibited unique advantages, improving mAP50 to 88.72% and mAP50:95 to 65.61% without increasing model complexity. This enhancement played a crucial role in optimizing bounding box regression precision and addressing sample imbalance issues while improving the model’s detection robustness across different target scales.

The synergistic combination of all three enhancement modules (Group 8) achieved optimal performance, with mAP50 reaching 94.23% and mAP50:95 rising to 72.52%, while maintaining a compact model size (2.65MB) and parameter count (0.94MB). Compared to the original YOLOv10 model, detection accuracy improved by 10.07 percentage points, while model size and parameter count reduced by 52.76% and 65.57%, respectively. These results conclusively demonstrate the synergistic effects of the enhancement modules in achieving both high-precision detection and model lightweighting objectives.

The ablation study results deeply reveal the core functions of each module: the MPIRU module serves as the foundation for lightweight design, significantly reducing model complexity; the c2f-DFAF module enhances multi-scale target detection capabilities through feature layer optimization; and the Focal-EIoU loss function optimizes detection accuracy while maintaining low computational overhead. The organic integration of these modules enables TomatoGuard-YOLO to achieve lightweight design while maintaining high performance, providing reliable technical support for resource-constrained practical agricultural application scenarios.

In the comparative experiments, we standardized the hyperparameters across all models to ensure a fair comparison. Specifically, all models were trained with the following hyperparameters: a batch size of 64, an initial learning rate of 0.001, the AdamW optimizer, and a cosine annealing learning rate scheduling strategy. Training was conducted for 300 epochs with an early stopping mechanism to prevent overfitting. Additionally, to maintain comparability, all experiments were performed on the same hardware environment (NVIDIA A100 GPU), which maximized the performance of each model while ensuring consistent experimental conditions. To assess the efficacy of the TomatoGuard-YOLO algorithm, we performed comprehensive comparative experiments alongside other state-of-the-art object detection algorithms. The results are summarized in Table 9 .

The comparative results presented in Table 9 clearly demonstrate that the TomatoGuard-YOLO model achieves a significant advantage in the mAP50 metric, reaching 94.23%, which is far superior to other compared algorithms. Especially compared to the original YOLOv10, TomatoGuard-YOLO improves by 10.07 percentage points, fully demonstrating the efficacy of our improvement strategies.

As for model complexity, TomatoGuard-YOLO’s parameter count is only 0.94M, a reduction of 65.57% compared to YOLOv10’s 2.73M. The model size is 2.65M, a reduction of 52.76% compared to YOLOv10’s 5.61M. This indicates that while improving detection accuracy, TomatoGuard-YOLO significantly reduces model complexity, making it a better choice for deployment in resource-limited environments.

Regarding inference speed, TomatoGuard-YOLO achieves an average frame rate of 129.64 FPS, an improvement of 3.20% compared to YOLOv10’s 125.62 FPS. This further demonstrates that despite the significant performance improvements, the model does not sacrifice inference speed; instead, it enhances inference efficiency. This high-efficiency and lightweight characteristic makes TomatoGuard-YOLO perform exceptionally well.

To further assess the benefits of TomatoGuard-YOLO, we performed a comparison of mAP and Loss curves with other models, as shown in Figures 11A, B .

From Figure 11A , the mAP curve for TomatoGuard-YOLO is notably superior to those of the other models under comparison, ultimately stabilizing at approximately 94.23%, which reinforces the model’s strong performance in tomato disease detection. Figure 11B illustrates that the initial loss for TomatoGuard-YOLO is 0.195359, significantly lower than that of the other models, eventually stabilizing at around 0.0626034. This suggests that the model achieves a quicker convergence and a lower final loss, further confirming its optimization efficacy. Overall, the outstanding performance of TomatoGuard-YOLO regarding accuracy, speed, and compact design highlights its significant potential for application in tomato disease detection within complex environments.

To thoroughly assess the TomatoGuard-YOLO model’s effectiveness in identifying different tomato diseases, we evaluated it using a custom imbalanced dataset. The detection results for various disease types, as achieved by the TomatoGuard-YOLO model, are summarized in Table 10 .

Table 10 demonstrates that the TomatoGuard-YOLO model attains notable Precision, Recall, and AP metrics above 90% for all four disease types and healthy samples, demonstrating high accuracy and recall rates. The model’s mAP reaches 94.23%, fully proving its excellent performance in handling different types of tomato diseases. Notably, the model excels not only in detecting common diseases such as late blight and early blight but also maintains high detection sensitivity for relatively rare diseases like gray mold and leaf mold. Additionally, the model’s outstanding performance in recognizing healthy samples helps reduce misdiagnosis and unnecessary treatments, which is crucial for practical agricultural production.

The AP50 values for detecting five types of tomato diseases and healthy samples using the proposed TomatoGuard-YOLO algorithm and other models are displayed in Table 11 . The findings clearly demonstrate that the TomatoGuard-YOLO algorithm shows enhanced adaptability in handling targets with pronounced sample imbalance and size variations.

As shown in Table 11 , the YOLOv10 algorithm achieves an AP50 of 90.71% for detecting healthy samples but only 77.18% for identifying leaf mold. Similarly, the YOLOv8n algorithm performs relatively well for late blight detection, with an AP50 of 87.16%, but struggles with gray mold, achieving only 78.53%. In contrast, the proposed TomatoGuard-YOLO algorithm demonstrates exceptional performance across all categories, significantly surpassing other models in the overall detection of five types of tomato diseases as well as healthy samples. Notably, for the more challenging cases of gray mold and leaf mold, TomatoGuard-YOLO achieves AP50 values of 93.07% and 90.36%, respectively, outperforming competing algorithms by a substantial margin.

Accurate identification and localization of tomato diseases are critical metrics for evaluating detection performance. While healthy samples and late blight are relatively easier to detect due to their distinct features, diseases such as early blight, gray mold, and leaf mold pose greater challenges. These diseases often exhibit subtle symptoms, with blurred boundaries between lesions and healthy tissue, making detection more difficult. Gray mold and leaf mold, in particular, are characterized by irregular lesion distribution, significant variations in lesion size, and high visual similarity to the background, further complicating accurate detection.

Compared to the next-best performer, YOLOv10, TomatoGuard-YOLO achieves a remarkable improvement of 10.08 percentage points in mAP50, with category-specific gains ranging from 6.88 to 13.18 percentage points. This comprehensive performance enhancement highlights the superiority of TomatoGuard-YOLO in tackling the complexities of tomato disease detection. Furthermore, it provides reliable technical support for precise diagnosis and timely intervention, offering significant practical value in real-world agricultural applications.

To visually compare the detection capabilities of the TomatoGuard-YOLO model, Figure 12 highlights the performance differences between YOLOv10 and TomatoGuard-YOLO in representative tomato disease scenarios.

The visual results clearly demonstrate that TomatoGuard-YOLO outperforms YOLOv10, particularly in handling complex backgrounds and multi-scale targets. These advantages underscore the robustness and precision of TomatoGuard-YOLO in accurately identifying and localizing diseased regions under challenging conditions, further validating its superiority in practical agricultural applications. In detecting small targets, TomatoGuard-YOLO demonstrates extremely high precision, especially in identifying early disease spots, accurately locating lesions. This is crucial for timely detection and control measures. In complex situations such as overlapping and occluded leaves, TomatoGuard-YOLO can still effectively distinguish disease areas, significantly reducing false detection rates, showing high robustness in complex scenarios. Additionally, in uneven lighting or complex shadow backgrounds, TomatoGuard-YOLO exhibits excellent environmental adaptability, accurately identifying disease areas, ensuring detection stability under different lighting conditions. Compared to YOLOv10, TomatoGuard-YOLO generates more precise bounding boxes, aiding in accurately assessing the severity and spread of diseases, providing more reliable support for actual disease assessment. Therefore, by introducing innovations such as MPIRU, DFAF, and Focal-EIoU loss functions, TomatoGuard-YOLO significantly improves detection accuracy and robustness while maintaining model lightweight, excelling in multi-scale and complex scenarios of tomato disease detection. Moreover, the model shows outstanding detection performance in quantitative evaluation metrics, providing strong technical support for early warning and precise control of tomato diseases in practical applications.

This research introduces a highly efficient and lightweight object detection approach built on an enhanced version of YOLOv10 for the challenging task of detecting tomato diseases—TomatoGuard-YOLO. By implementing an innovative model architecture and conducting thorough experimental evaluations, notable outcomes have been achieved. The key conclusions are as follows:

Experimental results indicate that TomatoGuard-YOLO achieves better performance than current methods on the tomato disease dataset. The model achieves an mAP50 of 94.23%, an improvement of 10.07 percentage points over the original YOLOv10, with a model size reduction of 52.76%, parameter count reduction of 65.31%, and an average inference speed increase to 129.64 FPS. These data fully demonstrate the method’s outstanding advantages in model accuracy, efficiency, and lightweight design.

Comparative experiments and visualization results further validate TomatoGuard-YOLO’s excellent performance in complex scenarios. Whether in small target detection, differentiation in complex backgrounds, multi-category disease recognition, or bounding box precision control, the proposed model shows significant improvement. These improvements not only provide reliable technical support for disease detection but also lay a solid foundation for applications in practical agricultural scenarios.

In summary, the TomatoGuard-YOLO model demonstrates excellent performance and broad adaptability in tomato disease detection tasks. Its high accuracy and reliability in disease detection provide powerful technical tools for disease prevention and intelligent management in tomato cultivation, offering significant application value in improving tomato cultivation efficiency and reducing disease losses.

Despite the significant achievements of TomatoGuard-YOLO in tomato disease detection, the current study lacks large-scale field deployment validation, which is crucial for understanding real-world performance. Future research can further explore the following aspects: