The dataset used in this study is derived from TomatoDiverse: An Open Dataset for Industrial Tomato Detection in Complex Natural Environments Tao (2024). The images were captured at tomato production bases located in Bayingolin Mongol Autonomous Prefecture, Xinjiang Uygur Autonomous Region, and Jingxian County, China. From this dataset, a total of 236 high-resolution images (4000×3000 pixels) depicting the growth posture of ripe tomatoes were selected.

These images reflect a variety of challenges present in complex natural environments, including variations in lighting conditions, different shooting distances, and occlusions. To enhance data diversity, the dataset was augmented using several techniques: random rotation (with angles randomly sampled within the range from -30 to +30) Cubuk et al. (2019), horizontal flipping Krizhevsky et al. (2012), and Gaussian blurring (with Gaussian kernel sizes randomly selected from [21, 31, 41, 51]) Zhang et al. (2019). After augmentation, the dataset was expanded to a total of 780 images. Figure 1 illustrates examples of the augmented images.

The criteria for determining red-ripe Choi et al. (1995) tomatoes in the dataset are as follows: 1. Visual Inspection: A tomato is labeled as ripe when the red area on its surface reaches or exceeds 90%. 2. Quantitative HSV-Based Determination: For samples where manual judgment is challenging, the Labelme tool is used to segment the tomato region, and the average values of Hue, Saturation, and Value (Brightness) are calculated in the HSV color space. A tomato is considered red-ripe if the following conditions are met: Hue (H): Within the range of 0 – 11.2 or 144 – 180 (where 11.2 is derived from dividing the visual threshold 10 by 0.9, and 144 is calculated similarly);Saturation (S): Between 100 and 255;Value (V): Between 100 and 255.

Mochen et al. (2023) adopted two keypoints (the peduncle and the picking point) to annotate and detect strawberries grown in an elevated cultivation system. In this cultivation mode, strawberry fruits hang naturally and maintain a stable posture, making keypoint annotation relatively easy. In contrast, tomatoes in natural growing environments may either hang down or lie horizontally on branches and leaves, exhibiting a wide variety of postures. To enable more effective localization of picking points, we introduce three keypoints in this study: P1 (picking point), P2 (peduncle location), and P3 (fruit bottom). P2 and P3 together provide information about the size and orientation of the fruit, which is critical for evaluating the feasibility of grasping. P1 is located approximately 1 – 2 cm above P2 along the direction of the peduncle and represents the ideal cutting point for the harvesting robot. This design ensures accurate picking without damaging the fruit itself and minimizes the risk of mistakenly cutting non-target parts such as branches or leaves.

Keypoint annotations for ripe tomatoes were performed using the Labelme software, with the tomato fruit and its picking points defined as annotation targets. Each ripe tomato was annotated using its minimum bounding rectangle, with the label assigned as “tomato”. Three keypoints were defined: P2, located at the junction of the calyx and the pedicel; P1, positioned 1 – 2 cm upward along the pedicel from P2; and P3, located at the bottom end of the tomato. An example of the annotated image is shown in Figure 2 . If a keypoint is partially occluded, the suffix ‘hide’ is appended to its name; if a keypoint is invisible or absent, the suffix ‘missing’ is used instead. All annotations are stored in standard JSON format, including the image path, image dimensions (width, height, number of channels), as well as the bounding box of the tomato and the coordinates of each keypoint. The JSON annotation files were converted into TXT format for training purposes. The dataset was split into training and testing sets at a ratio of 8:2, resulting in 624 images for training and 156 images for testing. The overall image distribution conforms to a 4:1 ratio between the training set and the testing set. The visibility statistics of the annotated keypoints are summarized in Table 1 .

YOLOv8-Pose is based on the YOLOv8 neural network and introduces a keypoint localization module for human pose estimation on top of object detection. It is primarily used for human detection and pose estimation. Both YOLOv8 and YOLOv5 models were developed by the same author. The network consists of five main components: the input, Backbone, Neck, Head, and output. The Backbone continues the structure of previous versions, where an input image of 640×640 is processed to generate feature maps of sizes 80, 40, and 20, corresponding to downscaling factors of 8, 16, and 32, respectively. The C2f module is extensively used, which, compared to the C2 module in YOLOv5, introduces deeper convolutional feature fusion, providing richer gradient information. The Neck follows the structure of YOLOv5, utilizing a path aggregation network to enhance feature fusion. It achieves this by upsampling low-level features and fusing them with high-level features, as well as downsampling high-level features and merging them with low-level features. This improves the network’s ability to handle objects at different scales. The output layer includes both classification and detection. The coupled head is replaced with a decoupled head, where classification and regression tasks are handled by separate branches. This separation improves efficiency and accuracy when handling different tasks.

YOLOv8-Pose is a keypoint prediction model based on human pose estimation. Compared to human pose estimation, the number of keypoints, target categories, and feature characteristics of ripe tomatoes differ significantly. Therefore, the original model is not directly applicable to the recognition of ripe tomatoes and the detection of picking keypoints, requiring improvements to YOLOv8-Pose. To address challenges such as target occlusion, leaf occlusion, and mis-segmentation caused by the similar color of ripe tomatoes, this study designs and introduces a Large Separable Kernel Attention (LSKA) module to enhance the network’s ability to distinguish target features Deng et al. (2024). Additionally, a weighted Bi-directional Feature Pyramid Network (BiFPN) is used to replace the FPN structure in the Neck, allowing the network to learn the importance of different input features through learnable weights, thereby improving feature fusion capabilities. The overall framework of the improved YOLOv8-Pose algorithm is shown in Figure 3 .

In object detection and keypoint estimation tasks, effectively acquiring and processing multi-scale feature information remains a major challenge. The traditional Feature Pyramid Network (FPN) ( Figure 4a ) Liu et al. (2018) aggregates multi-scale features in a top-down manner, which is often constrained by unidirectional information flow. To address this, the Path Aggregation Network (PANet) Tan et al. (2020) introduces an additional bottom-up path aggregation, as illustrated in Figure 4b . In the Neck of YOLOv8-Pose, the FPN structure from YOLOv5 is still adopted. High-level feature information is fused through the FPN+PAN structure, which improves the efficiency of feature transmission. However, this also increases computational complexity. Particularly when dealing with high-resolution inputs, it may result in high computational costs and reduced real-time performance. Moreover, the fixed architecture lacks adaptive adjustment capabilities for different tasks and datasets. To address the aforementioned issues such as unidirectional information flow, high accuracy at the cost of numerous parameters and heavy computation, as well as information loss and redundancy caused by simple feature concatenation, this paper proposes replacing the original FPN+PAN structure in YOLOv8-Pose with a BiFPN. Based on PANet and NAS-FPN ( Figure 4c ), BiFPN optimizes multi-scale feature fusion strategies. Its architecture is illustrated in Figure 4d .

Different feature input resolutions contribute differently to the final feature network output. Therefore, the network needs to learn the corresponding weights to balance feature importance. As shown in Figure 4d , feature maps from different levels are input into the BiFPN_Concat layer (taking BiFPN_Concat2 as an example). Given two feature maps F 0 ∈ R C 0 × H × W and F 1 ∈ R C 1 × H × W , where C ₀ and C ₁ denote the number of input channels, H and W represent the height and width of the feature maps respectively, the weights in the BiFPN_Concat2 module are initialized as w=[1,1], assigning equal initial importance to each input feature map. During each forward propagation step, weights are normalized to ensure stable training. To prevent division by zero, a small constant ϵ = 0.0001 is used. The formula for weight normalization is as follows (Equation 1):

In BiFPN_Concat2, the number of input feature maps is n = 2. After normalization, the weights are represented as w e i g h t = [ w e i g h t 0 , w e i g h t 1 ], where w e i g h t 0 + w e i g h t 1 = 1. Each input feature map is multiplied by its corresponding normalized weight: F 0 w e i g h t e d = w e i g h t   0 × F   0 , F 1 w e i g h t e d = w e i g h t   1 × F   1 During training, the model optimizes w based on loss feedback, which in turn updates weight values dynamically. Finally, BiFPN_Concat2 concatenates the weighted feature maps along the channel dimension to obtain the output feature map (Equation 2):

At this stage, the output feature map is represented as F o u t = R ( C 0 + C 1 ) × H × W , successfully fusing feature information from F ₀ and F ₁.

To address the challenges of occlusion among small targets and mis-segmentation caused by similar colors between ripe tomato fruits, this paper designs and incorporates a Large Separable Kernel Attention (LSKA) module to enhance the network’s ability to distinguish target features Deng et al. (2024). By enlarging the receptive field of the convolutional kernels, the module focuses more on the overall shape characteristics of the target, thereby reducing the interference from background textures and color similarity in the detection results Ding et al. (2022). As the kernel size increases, LSKA can effectively capture broader contextual information, thus improving the recognition performance of small targets in complex scenes. However, directly increasing the kernel size also significantly raises the computational cost and memory usage, posing challenges to the model’s efficiency. LSKA decomposes a k × k depthwise convolution (DW-Conv) into two sequential 1D DW-Convs with kernel sizes of 1×(2d −1) and (2d −1)×1, as well as two sequential 1D depthwise-dilated convolutions (DW-D-Conv) with kernel sizes of ⌊ k d ⌋   ×   1 and 1 × ⌊ k d ⌋ . The outputs are then passed through a 1×1 convolution. The depthwise-dilated convolutions are responsible for capturing global spatial information from the depthwise convolution output. This design achieves an effective receptive field equivalent to that of a large kernel while avoiding the quadratic increase in computational cost typically caused by using large kernels in depthwise convolutions alone. Given an input feature map F ₀ ∈ R ^C×H×W, where C is the number of input channels and H and W represent the height and width of the feature map, the output of LSKA is computed as follows (Equations 3–6):

Where d is the dilation rate and k represents the maximum receptive field. * and ⊗ denote convolution and Hadamard product (element-wise multiplication), respectively. Z C ¯ represents the output of two cascaded 1D depthwise convolutions (DW-Conv) with kernel sizes of 1 × ( 2 d − 1 ) and ( 2 d − 1 ) × 1 , respectively. Z C represents the output of two cascaded 1D depthwise dilated convolutions (DW-D-Conv) with kernel sizes of ⌊ k d ⌋ × 1 and 1 × ⌊ k d ⌋ , respectively (where ⌊ . ⌋ denotes the floor function). A C represents the output of applying a 1 × 1 convolution to Z C . F C ¯ represents the result of the Hadamard product between the feature map A C and the input feature map F C

The LSKA attention mechanism decomposes large convolutional kernels to achieve a larger receptive field, enhancing the model’s feature extraction capability. We set kernel sizes of 7, 11, 23, 35, 41, and 53, and trained the YOLOv8-LBP model on a self-collected red-ripe tomato dataset. When the kernel size was 23, the model achieved the highest Precision of 0.944; at kernel size 7, the mAP ₅₀ peaked at 0.926 with the smallest parameter count of 3.164M; kernel size 41 yielded the best mAP ₅₀₋₉₅ and keypoint mAP ₅₀₋₉₅ − kp scores of 0.71 and 0.785, outperforming other sizes; while kernel size 53 resulted in the highest FPS of 107.53. Considering the balance between detection performance and computational efficiency, we selected the LSKA attention mechanism with kernel size 41 for subsequent experiments. Details are shown in Table 3 .

In the experiments on object recognition and keypoint prediction, we evaluated the performance of several models including YOLOv5-Pose, YOLOv8-Pose, YOLOv11-Pose, YOLOv12-Pose, YOLOv8-Pose-ss Guo et al. (2024), YOLOv8-Pose-cv Mochen et al. (2023), and YOLOv8-LBP on a self-collected red-ripe tomato dataset. The results are presented in Table 4 . Comparative analysis shows that YOLOv8-LBP achieved 92.7%, 91.7%, 71.%, and 78.5% on Precision, mAP ₅₀, mAP ₅₀₋₉₅, and mAP ₅₀₋₉₅ − kp metrics, respectively. Although its Precision is slightly lower than YOLOv8-Pose-cv’s 0.945, it surpasses all other models in the remaining metrics, demonstrating its outstanding performance in red-ripe tomato detection and harvesting keypoint localization tasks. Figure 5 illustrate the experimental performance and comparative results of different models across various metrics.

The experimental results comparing YOLOv5-Pose, YOLOv8-Pose, YOLOv11-Pose, YOLOv12-Pose, YOLOv8-Pose-ss, YOLOv8-Pose-cv, and YOLOv8-LBP models for ripe tomato recognition and picking keypoint detection are shown in Figures 6 , 7 . From column (i), it can be seen that models b and c have certain deviations in keypoint prediction, while models d, e, f, and g have lower tomato recognition confidence compared to model h; Column (ii) shows that under leaf occlusion, models c, d, e, and f exhibit multiple slight deviations in keypoint prediction, models b and g show larger deviations, while model f, despite some minor deviations, achieves better overall keypoint prediction accuracy than other models; Column (iii) indicates that models c, d, e, and g have false positives in tomato recognition, models c, d, e, f, and g all show certain deviations in keypoint prediction, with model f having the most accurate keypoint predictions and the highest tomato recognition confidence.

Figures 8 , 9 present a comparison of different models’ results on object recognition and keypoint detection in complex natural environments. Column (i) includes cases with hand background interference, clusters of tomatoes, and environmental occlusions; column (ii) shows scenarios with strong direct sunlight; column (iii) shows tomato clusters and fruit occlusions. Yellow arrows indicate false positives, while blue arrows denote other types of errors. In column (i), models c, d, e, and f exhibit three or more false positives; models b, g, and h have two false positives; models c, d, e, and g show missed detections of tomatoes; models b, f, and h show no missed detections; models b, d, e, and g have large deviations in keypoint predictions within multiple bounding boxes; models c, f, and h show large deviations within a single bounding box. In column (ii), models b, c, d, e, f, and g all have false positives under shadow conditions; models b, c, e, and g have false positives under direct sunlight; model g has two bounding boxes with large keypoint prediction deviations, while the other models have one such deviation each; model h has no false positives. In column (iii), models d, e, and f have false positives; models c, d, e, and f have missed detections; model g shows repeated detection of tomatoes; model h accurately recognizes every ripe tomato; models b, c, g, and h have large keypoint prediction deviations. Overall, model h outperforms the others. YOLOv8-LBP effectively addresses missegmentation caused by similar tomato fruit colors and occlusions in natural environments, providing more accurate keypoint localization, making it suitable for robotic harvesting applications.

To comprehensively evaluate the performance of different models, three commonly used computational metrics were selected for comparative analysis: the number of parameters (Params), Giga Floating-point Operations per Second (GFLOPs), and frames per second (FPS). The number of parameters (Params) represents the total count of all trainable parameters in the model, reflecting its complexity and storage requirements. GFLOPs measure the computational cost required for a single forward inference. FPS indicates how many image frames the model can process per second during actual operation and is an important metric for evaluating real-time performance. Table 5 presents a comparison of these three computational metrics across different models. As shown in Table 5 , YOLOv8-Pose-ss has the smallest number of parameters, the highest FPS, and the lowest computational cost (GFLOPs), indicating that our proposed model does not have an advantage in computational efficiency. Compared with the baseline model, our approach slightly increases the number of parameters (from 3.08M to 3.175M), raises GFLOPs by 0.1, and marginally improves FPS (from 96.15 to 99.01). Although computational cost increases slightly, this overhead is reasonable and acceptable, especially given the substantial improvements in accuracy: relative to the baseline model, object detection Precision improves by 4.5%, mAP ₅₀ by 1.1%, mAP ₅₀₋₉₅ by 2.8%, and keypoint detection accuracy (mAP ₅₀₋₉₅ − kp) by 3.3%. Compared with YOLOv8-Pose-ss, our model achieves 4.7% higher Precision (0.927 vs. 0.880), 2.3% higher mAP ₅₀ (0.917 vs. 0.894), 6.3% higher mAP ₅₀₋₉₅ (0.710 vs. 0.647), and 12.2% higher mAP ₅₀₋₉₅ − kp (0.785 vs. 0.663). These results indicate that our approach achieves a favorable trade-off between accuracy and computational cost. Furthermore, with the continuous advancement of embedded device computing power and the widespread adoption of cloud computing, the impact of a slight increase in computational resources on practical deployment is gradually diminishing. Therefore, we believe the proposed method retains strong practical applicability and scalability while maintaining high accuracy.

To verify the effectiveness of each module, an ablation study was conducted based on the YOLOv8n-Pose model by gradually adding and replacing components. The results are shown in Table 6 . Compared to the original YOLOv8n-Pose, adding the LSKA attention mechanism alone expanded the model’s effective receptive field, enabling the extraction of richer feature information. This led to increases of 0.7% and 0.5% in mAP ₅₀ and mAP ₅₀₋₉₅, respectively. When adding the BiFPN module alone, the model’s ability to process multi-scale feature information was enhanced, resulting in a 1.9% improvement in precision (P). When both LSKA and BiFPN were added simultaneously, the model not only obtained richer feature information but also leveraged BiFPN’s ability to adaptively adjust the importance of these features, achieving more effective feature fusion. This allowed the model to better focus on the overall shape features of the target while reducing the interference of background textures and color similarity, leading to improvements of 4.5%, 1.1%, 2.8%, and 3.3% in P, mAP ₅₀, mAP ₅₀₋₉₅, and mAP ₅₀₋₉₅−kp, respectively. The ablation experiments demonstrate that the two improvements proposed in this study are effective.