Optimize the network of YOLOv8 for grape detection task

The aim of the fruit and vegetable object detection task is to enable the picking robots to identify objects efficiently and accurately. Although there are extensive studies in improving the accuracy and speed of fruit and vegetable object detection task, the methods still lack of responsiveness to different requirements of picking robots and algorithms for picking different objects. In this paper, based on YOLOv8, Lightweighting Sequential Polarized Self-Attention YOLOv8s (LSPA_YOLOv8s) was proposed to improve the accuracy and speed in grape object detection task. The new approach has three critical designs: (1) Incorporating the ResNet and multiple attention mechanisms to improve the PSA, the DSPSA module is proposed to improve feature extraction; (2) Replace the convolution block in C2f with PConv to get the optimized C2f_PConv module, take advantage of the nature of partial convolution to cut the number of model parameters and memory accesses; (3) Replacement of some Conv blocks in YOLOv8s with Involution to further increase network speed. Experiments show that the method performs well on the grape object detection task.