An improved algorithm based on YOLOv5s was proposed in order to accurately identify different fruit targets on fruit trees and solve problems such as unsatisfactory fruit recognition accuracy. Firstly, the EIoU loss function is applied to YOLOv5s to minimize the difference between the width and height of the target frame and the anchor frame. Secondly, the ECA-Net attention mechanism was added to the backbone network to better extract the characteristics of different fruit targets. The experimental results show that YOLOv5s+ECA+EIoU loss has obvious comprehensive performance advantages. Compared with several existing network algorithms, various data of the improved algorithm is superior to other algorithms, thus verifying the effectiveness of the improved algorithm and providing necessary technical support for the research of object detection during robot picking.