YOLOv8-LBP: Multi-Scale Attention Enhanced YOLOv8 for Ripe Tomato Detection and Harvesting Keypoint Localization

Yong Wang^*Shunfa XuKonghao Cheng

• School of Computer Science and Technology., Zhejiang University of Technology, Hangzhou, China

In the process of target detection for tomato harvesting robots, there are two primary challenges. First, most existing tomato harvesting robots are limited to fruit detection and recognition, lacking the capability to locate harvesting keypoints. As a result, they cannot be directly applied to the harvesting of ripe tomatoes. Second, variations in lighting conditions in natural environments, occlusions between tomatoes, and missegmentation caused by similar fruit colors often lead to keypoint localization errors during harvesting. To address these issues, we propose YOLOv8-LBP, an enhanced model based on YOLOv8-Pose, designed for both ripe tomato recognition and harvesting keypoint detection. Specifically, we introduce a Large Separable Kernel Attention (LSKA) module into the backbone network, which effectively decomposes large kernel convolutions to extract target feature matrices more efficiently, enhancing the model's adaptability and accuracy for multi-scale objects. Secondly, the weighted bidirectional feature pyramid network (BiFPN) introduces additional weights to learn the importance of different input features. Through top-down and bottom-up bidirectional paths, the model repeatedly fuses multi-scale features, thereby enhancing its ability to detect objects at multiple scales. Ablation experiments demonstrate that, on our self-constructed ripe tomato dataset, the YOLOv8-LBP model achieves improvements of 4.5% in Precision (P), 1.1% in mAP50, 2.8% in mAP50−95, and 3.3% in mAP50−95 −kp compared to the baseline. When compared with the state-of-the-art YOLOv12-Pose, YOLOv8-LBP shows respective improvements of 5.7%, 0.5%, 3.5%, and 4.9% in the same metrics. While maintaining the improvement in model accuracy, our method introduces only a small computational overhead, with the number of parameters increasing from 3.08M to 3.175M, GFLOPs rising by 0.1, and the inference speed improving from 96.15 FPS to 99.01 FPS. This computational cost is reasonable and acceptable. Overall, the proposed YOLOv8-LBP model demonstrates significant advantages in recognizing ripe tomatoes and detecting harvesting keypoints under complex scenarios, offering a solid theoretical foundation for the advancement of robotic harvesting technologies.