Existing tomato datasets primarily consist of images featuring single or multiple tomatoes against relatively clean and unobstructed backgrounds, limiting their applicability to real-world field production scenarios. Therefore, this study collected data from a small tomato cultivation base located at the International Internet Agricultural Expo Park in Wuzhen City, Zhejiang Province, where tomatoes at various growth stages were cultivated for market supply. Data acquisition was conducted over the period from April 3 to May 30, 2024. A SCOUT 2.0 robot equipped with an iPhone 14 Pro mounted horizontally was utilized to capture images of the small tomato plants. A total of 160 plants were arranged in two rows, each extending 20 meters in length. During data collection, the robot moved at a constant speed, photographing each row sequentially from left to right and then returning from right to left to capture images in the opposite direction. The acquired images were subsequently uploaded to a PC for further processing. The overall experimental setup is illustrated in Figure 1 , with the right panel depicting the robot in operation.

The photographed images were standardized to 640*640 pixels, labeled using the LabelImg tool, with the following labels: green fruit tomato, red fruit tomato and yellow fruit tomato. Following agronomic standards and harvesting requirements, these three labels correspond to unripe tomatoes, ripe tomatoes, and tomatoes between unripe and ripe stages, respectively. Such three classifications can fit the actual production decisions and reduce redundant judgements, while ensuring the efficiency of data annotation and data processing (Wan et al., 2018). Finally, the original small tomato data samples were obtained as 661, which were divided into training set, validation set and test set according to the ratio of 3:1:1. Given that the training set consists of only 535 images with only a small portion of overexposure and blurring under natural light conditions, in order to enhance the generalization of the model, six data enhancement techniques were adopted to process the training set data, including exposure, rotation, blurring, random brightness adjustment, mirroring and noise addition. As shown in Figure 2 , each image is enhanced by taking a random combination of three of the above enhancement methods.

A notable feature of this dataset is the inclusion of complex distractions from real environments, with varying degrees of occlusion problems on each image. Based on the type of blocked, we grouped the detection difficulties into four categories: extreme tiny tomatoes, mutual shading between classes, branch stalk shading, and leaf shading. The proportion of tiny small tomatoes was the largest, with a more similar amount of interclass shading and leaf shading, and a relatively small amount of branch and stem shading.(The number of different maturity categories in the original dataset and the corresponding number for each shading type are shown in Table 1 ) In order to mitigate the impact that category imbalance would have on training, we used the CopyPaste method to perform an additional data augmentation operation on red and yellow fruit tomatoes, which is to copy the instances in the image containing red and yellow fruit tomatoes and paste them into another image during the training process, adding instances from fewer categories to generate new training samples.

Combined with the growth characteristics of the small tomato itself, during the fruiting period, the growth height of the small tomato ranges from 0.5 to 2.3 meters, and within a relatively short period of time, there is a large span of height change. In order to be able to meet the normal work at different heights, we have also designed and improved the agricultural robot hardware accordingly. Firstly, the robot stand is built by 1.5mm iron plate, and the overall structure is in the shape of a tower, which is divided into three layers to meet the needs of different sensors and different heights of mounting. Next, electrical adapter devices are fixed on the bottom layer for powering the sensors of each device and edge computing devices are fixed on the bottom layer for processing real-time data. To improve the stability of the collected data, the camera head is mounted on the bottom tail, and the shooting camera is mounted on the camera head tilt rotation connector. The middle layer installs the router used by the robot for communication, which facilitates remote operation and control of the robot. The top layer is fitted with LIDAR to prevent other equipment from interfering with the laser. In this work, we deployed the detection algorithm ultimately on an edge device and utilized an agricultural robot to achieve the work of detecting and counting small tomatoes of different ripeness in a facility greenhouse, overcoming the problem of occlusion during the growing process. The detailed hardware composition as well as the field applications are shown in Figure 3 .

In real production environments, the large-scale planting of small tomatoes has significant advantages in improving production efficiency, but in the growth process of small tomatoes, regularized planting makes the lush branches and leaves obscure the fruits, and changes in the intensity and angle of the sunlight at different moments also significantly change the brightness and contrast of the image, making it more difficult to count the fruits.

This study proposed a Ta-YOLO model for the detection and counting of small tomatoes in a real production environment to address these challenges. The model retained the overall framework of YOLOv8n, adopted C2f-rghost combined with Space-to-Depth Conv module to reconstruct the backbone structure, and at the same time, the DASPPF structure was proposed to enhance the fine-grained representation. And the CSAM multiple attention mechanism was created in the neck structure, and an additional detection head was added to enhance the detection ability in different scales and occlusion situations. Finally, the EWDIoU loss function was proposed to improve the detection accuracy for small tomatoes. The overall structure of Ta-YOLO is shown in Figure 4 These improvements will be further illustrated above.

To expand the training samples and enhance the model’s generalization ability, robustness, and adaptability in real-world applications, we apply data augmentation techniques such as exposure adjustment, rotation, blurring, random brightness variation, and mirroring to simulate diverse scene variations. The experimental results are presented in Table 3 .

After data augmentation, the overall mAP@50 showed a noticeable improvement, indicating enhanced detection performance of the model. In addition, the precision values for all categories and the overall precision increased, suggesting a reduction in false positives and more effective feature learning. While the recall of the ytomato class improved, the recall of other categories slightly declined. This is likely due to increased background complexity in the augmented images, which made the model more conservative in its predictions, leading to a higher miss rate for true targets. However, since the ytomato class had relatively few samples before augmentation, the augmented data effectively alleviated the issue of data scarcity and helped the model learn more stable features. Although the mAP of gtomato and rtomato slightly decreased—possibly due to distributional shifts or reduced feature stability caused by augmentation—the overall improvement in mAP indicates that the model became more balanced and achieved better generalization.

In this section, the fully enhanced model is compared with simplified variants incorporating individual improvements to independently evaluate the effectiveness of each component. All enhancements are built upon the baseline YOLOv8n model, targeting the following aspects: modifications to the YOLOv8n backbone; introduction of a novel feature pyramid pooling structure, DASPPF; incorporation of a lightweight detection head; integration of the CSAM attention mechanism to improve multi-scale feature fusion; and replacement of the traditional loss function with the proposed EWDIoU loss, which leverages a two-dimensional Gaussian distribution to enhance bounding box regression. In Table 4 , each improvement was incrementally incorporated into the baseline model, and the corresponding performance metrics were evaluated. Specifically, “A” denotes the backbone enhancement, “B” refers to the proposed DASPPF module, “C” indicates the addition of a small object detection head, and “D” represents the proposed CSAM module. The final model, Ta-YOLO, integrates all these enhancements.

The results of the comprehensive ablation study are summarized in the table, highlighting the following key findings: (1) Lightweight modifications to the backbone successfully reduced the parameter count and increased inference speed (FPS), albeit at the cost of reduced accuracy. (2) The proposed DASPPF feature pyramid pooling structure significantly enhanced the extraction of salient features, with recall rates for green and red tomatoes reaching 70.1% and 82.4%, respectively—improvements of 1.4% and 4.3% over the baseline. Furthermore, the mAP increased by 0.5% and 5.8% compared to the baseline. These results indicate that preserving global contextual information in complex real-world scenes facilitates more accurate target recognition. (3) The addition of the tiny detection head increased the mAP to 83.3%, while simultaneously reducing the parameter count and improving FPS relative to the baseline. However, this enhancement resulted in an increased computational load. These findings indicate an improved multi-scale detection capability, rendering the model more effective for small tomato detection. (4) The CSAM attention mechanism further enhanced recognition accuracy by efficiently integrating multidimensional feature information, particularly benefiting the detection of multiple small tomatoes at image edges or under occlusion. Moreover, the proposed EWDIoU loss function effectively addressed challenges associated with small target detection, yielding superior performance across small tomato categories. Across all evaluated samples, the mAP for heavily shaded green and red tomatoes improved from 78.1% and 82.2% to 81.0% and 87.2%, respectively, demonstrating the targeted effectiveness of our approach in mitigating shading-related challenges. Furthermore, a comparative analysis between the original baseline and Ta-YOLO under complex real-world conditions, including occlusion, is presented in Figure 9 . The results confirm that Ta-YOLO achieves superior detection performance in these challenging scenarios.

The two-stage detection algorithm initially extracts candidate regions from the input image, followed by classification of each candidate. To evaluate the effectiveness of the proposed method, it was compared against the classical two-stage detector Faster R-CNN. Additionally, the enhanced model was benchmarked against several widely used one-stage detection algorithms, including YOLOv5, YOLOv7, YOLOv9, YOLOv11, and HyperYOLO. To ensure fairness and emphasize the effectiveness of the proposed method, comparisons were performed using lightweight variants of the evaluated algorithms. The experimental results, presented in Table 5 , indicate that Ta-YOLO achieved recall and mAP values of 76.9% and 84.4%, respectively, outperforming most competing models. Significantly, YOLOv9s attained the highest mAP of 85.3% in this comparison. However, YOLOv9s exhibited a parameter count of 37.88 million and a computational complexity of 40.6 GFLOPs, exceeding those of Ta-YOLO by 27.3 million parameters and 26.1 GFLOPs, respectively. Additionally, YOLOv9s demonstrated lower FPS performance compared to Ta-YOLO, underscoring the trade-off between accuracy and computational efficiency. Although YOLOv7’s mAP was only 0.4% lower than that of Ta-YOLO, its parameter count was approximately thirteen times greater. Compared to YOLOv11 and HyperYOLO, Ta-YOLO achieves higher FPS with a similar parameter count, while demonstrating superior accuracy.

Nevertheless, examination of the table reveals that, despite Ta-YOLO’s superior overall performance compared to other detectors, its recall for green tomatoes is below the average recall, indicating the presence of false negatives in green tomato detection. Moreover, this issue is not unique to Ta-YOLO but is prevalent across most detection models. An analysis of the dataset revealed that extensive leaf shading on green tomatoes contributes to erroneous detections. Notably, the dataset was annotated with stringent criteria, including labeling tomatoes even when heavily occluded by foliage, which may further contribute to the detection challenges observed. It is worth noting that, YOLOv9 achieves a relatively higher recall for green tomatoes. Our analysis attributes this to YOLOv9’s heavier parameterization, which facilitates more precise alignment of feature map edges. Consequently, future work will focus on enhancing edge and texture perception by improving the extraction and representation of edge features.

In this section, four representative challenging cases from the dataset were selected, as illustrated in Figures 10A, D, G, J . In these figures, yellow circles denote missed detections, blue circles indicate false positives, and orange squares mark regions with increased identification difficulty. In Figure 10B, E , the leaves highlighted by blue circles were erroneously classified as green and red tomatoes, respectively. In comparison, the proposed algorithm correctly avoids these misclassifications, as demonstrated by the absence of false positives within the blue dashed circles in Figure 10C, F . In Figure 10H , occlusion caused by tomato branches adversely affects detection, resulting in the tomato marked by the blue circle being erroneously identified as multiple instances. In contrast, the corresponding region within the blue dashed circle in Figure 10I is correctly detected by the proposed method. Likewise, in Figure 10J , extensive occlusion from the tomato petiole leads to a missed detection of the small tomato indicated by the yellow circle in Figure 10K , whereas Figure 10L shows successful recognition. The same four challenging cases presented in Figure 10 are used to visualize and compare the detection results between YOLOv8 and Ta-YOLO in Figures 11A, D, G, J . In Figure 11B , the very small tomatoes indicated by yellow circles were completely missed, whereas those within the yellow dashed circles in Figure 11C were accurately detected, including occluded instances. In Figure 11E , the tomato enclosed by the yellow circle was heavily obscured and not correctly identified; however, the improved algorithm presented in this study successfully detected it in Figure 11F . Similarly, Figure 11H exhibits the same issue observed in Figure 10H , where the tomato marked by the blue circle was mistakenly identified as multiple instances, whereas the corresponding region in Figure 11I within the blue dashed circle was correctly recognized. In Figure 11J , occlusion caused by the tomato petiole led to a missed detection of the small tomato marked by the yellow circle in Figure 11K , while Figure 11L demonstrates its accurate detection. Collectively, these results demonstrate that Ta-YOLO achieves higher accuracy and greater robustness in detecting shaded small tomatoes under real production conditions.

In deep learning, the attention mechanism, which simulates the selective focus of human cognition, has been extensively applied across domains including image processing and natural language processing. Among various attention mechanisms, the Global Attention Module (GAM) stands out as a global attentional approach that effectively preserves the majority of salient information, thereby enhancing feature interaction. The SE attention mechanism emphasizes effective feature extraction by employing a squeeze-and-excitation process that encourages the network to integrate spatial and channel information within the local receptive field. Differently, the scSE attention mechanism simultaneously combines spatial and channel attention in parallel to enhance feature representation. Similarly, the CBAM attention mechanism facilitates feature interaction through sequential fusion of channel and spatial attention. Each of these attention mechanisms, having demonstrated strong performance across various tasks, was integrated into the Ta-YOLO model for comparative evaluation against CSAM. The results showed that, CSAM achieved the highest accuracy, with mAP and recall values of 84.4% and 76.9%, respectively. Among the competing mechanisms, CBAM exhibited the lowest parameter count and computational cost, with 9.62 million parameters and 14.2 GFLOPs, although its mAP and recall were 1% and 2% lower than those of CSAM. GAM, despite having the largest parameter count at 15.88 million, attained an mAP of 83.4%, comparable to CBAM, thus neither surpassing CSAM’s performance nor justifying the increased complexity. Additionally, CSAM maintains the same plug-and-play compatibility as these established attention modules. Table 6 presents the detection performance of Ta-YOLO on the real tomato dataset, while Figure 12 illustrates heatmap visualizations corresponding to different attention mechanisms. The results clearly indicated that the proposed CSAM module outperforms others by effectively concentrating on heavily occluded and small-sized tomatoes. In summary, the integration of spatial and channel attention within the CSAM module yields superior detection efficacy.

The loss function plays a critical role in object detection tasks by quantifying the discrepancy between model predictions and ground truth labels. This measurement guides model optimization during training, ultimately improving detection performance. In this study, several widely used loss functions were evaluated on the Ta-YOLO model and compared against the proposed EWDIoU loss function to assess its effectiveness. The corresponding experimental results were summarized in Table 7 . Among the evaluated loss functions, the EWDIoU achieved the highest mAP@50 and recall scores of 84.4% and 76.9%, respectively. Although the GIoU loss attained an identical recall value, its accuracy was lower at 83.0%, representing a 1.4% deficit compared to EWDIoU. The EIoU loss reached a mAP@50 of 83.8%, close to the highest value; however, its precision and recall metrics were inferior to those of the EWDIoU loss function. Although the original CIoU loss used in YOLOv8 achieved the highest precision of 88.2%, its mAP was limited to 82.3%, and recall reached only 75.1%. These results indicated that the model struggled to detect all valid targets, particularly small tomatoes with occluded edges, leading to missed detections. Consequently, this shortcoming contributed to the decline in both recall and mAP. The EWDIoU loss function employs a two-dimensional Gaussian distribution approach to process discretized data, effectively addressing the bounding box insensitivity to small tomato targets and thereby enhancing detection accuracy. The proposed improvements demonstrate notable performance gains, indicating that targeted enhancements can overcome common detection challenges, including performance degradation in complex environments characterized by occlusion and small targets, as well as reducing instances of misdetection and omission.

In the proposed EWDIoU loss function, to effectively mitigate the impact of small target tomato bounding boxes on the loss calculation while preserving the detection performance advantages for larger target tomatoes, an adjustable hyperparameter λ was introduced. This hyperparameter balances the contribution of the IoU in the loss function calculation, allowing for adaptive adjustments across different target scales. On the one hand, λ suppresses the bias amplification effect caused by the smaller scale of small targets in bounding box errors. On the other hand, it ensures the importance of large targets is preserved in the detection task, thereby achieving a dynamic balance and optimizing the loss function’s performance for detecting targets of varying scales. The model’s robustness and accuracy in handling multi-scale targets were significantly enhanced. We conducted experiments with 10% intervals and keeping the criterion of λ 1 + λ 2 = 1 to observe the effects brought by different values on the overall detection results, as shown in Figure 13 . The number of experiments is 9 groups in total, and the experimental results showed that when λ 1 is 0.7 and λ 2 is 0.3, our EWDIoU effect performed the best on the small tomato dataset of the field scene, and its total category mAP value reached 84.4%. In the comparison across different categories, the mAP trends of green-fruited tomatoes and yellow-fruited tomatoes exhibited high consistency with the overall category mAP. With the highest mAP in the total category, green-fruited tomatoes and yellow-fruited tomatoes reached 81% vs. 84.9% mAP, respectively. Our data analysis showed that green-fruited tomatoes had a higher probability of being obscured in the sample of obscured small tomatoes, and that the number of green-fruited tomatoes in the sample of small targets was relatively large. Therefore, when the mAP of green-fruited tomato reached the optimum, the mAP of the total category also reached the peak, which further validated that the EWDIoU loss function is able to effectively solve the problem of the detection of occluded fruits. In addition, in the practical application scenario, the shading rate of yellow-fruited tomato was much lower than that of green-fruited tomato, but its mAP was still able to keep the same trend with the total category mAP, which indicated that the loss function has strong robustness and adaptability. The change in mAP of red-fruited tomato was relatively smooth, and the fluctuation range of its mAP was controlled within 2.1%. This phenomenon can be primarily attributed to the relatively moderate shading in red-fruited tomato samples, coupled with greater color variability arising from differing ripeness levels. Nonetheless, the consistent detection performance suggests that the EWDIoU loss function does not induce significant errors or overfitting when applied to this category. In summary, EWDIoU demonstrates strong adaptability and stability across varying degrees of occlusion scenarios.