SDSLP-YOLO: An End-to-End Model for Simultaneously Detecting Strawberry and Locating Picking Point

Accurately identifying strawberries and locating picking points are prerequisites for harvesting robots. To address the issues of multiple detections, missed detections, and false detections, as well as inaccurate picking point localization caused by large differences in fruit shape, complex backgrounds, and occlusions during the picking process, this paper proposes a method based on the YOLOv8-pose model. First, a dual-path feature extraction backbone network (LGFBNet) that combines convolution and Transformer is proposed. This module enhances the richness and expressiveness of features by coupling global and local features, effectively solving the issue of multiple detections caused by large differences in fruit shape. Secondly, CBAM is introduced into the neck of the network, allowing the network to focus more on the distinguishing features between the strawberries and the background, effectively addressing the missed and false detections caused by complex backgrounds. Finally, the NWD-CIoU weighted loss function is introduced, where the scale-balancing property of NWD compensates for the sensitivity of CIoU to positional shifts, reducing the additional missed detection rate caused by severe occlusions. Experimental results show that the proposed method achieves an average precision (mAP-obj) of 88.3% for strawberry detection and an average precision (mAP-kp) of 74.2% for keypoint detection, which are 1.7% and 3% higher than YOLOv8-pose, respectively. In summary, this method has significant theoretical and practical implications for tasks such as strawberry recognition and picking point localization in complex environments.