Research on Pear Detection Algorithm Based on Improved YOLOv7

In pear-picking, the method of pear detection directly affects the accuracy and efficiency of the picking process. To address the issue of low detection accuracy in orchards, this paper proposes the ViT-YOLOv7 pear detection model. First, a backbone network based on the Vision Transformer is designed to extract features of the fruits in complex environments using convolution and self-attention mechanisms. Next, a lightweight feature fusion network is constructed by introducing partial convolution to replace traditional convolution operations. Finally, to address the issues of misidentification and missed detection caused by fruit occlusion, a decoupled detection head is designed, and a no-anchor strategy is adopted for fruit detection. Experimental results show that compared to YOLOv7-Tiny, the ViT-YOLOv7 model can accurately identify and locate pears with high efficiency, ensuring that picking operations are performed quickly and precisely, thereby improving overall picking efficiency.