Research on compression method of yolov5 model based on channel pruning

With the development of high-resolution and wide -format SAR satellites, on-board real-time processing of massive data has become a major development trend. However, due to the limitation of on-board resources, on-board real-time processing is faced with insufficient storage space, insufficient operating memory, and a large amount of computation. In order to solve this problem, this paper conducts channel pruning for the large target detection model yolov5 to achieve model compression, uses the bn layer as the basis for pruning, and adaptively determines the pruning ratio of each layer on the basis of analyzing the pruning sensitivity. We experimentally determine the best retraining strategy. In the end, this paper obtained the yolov5 lightweight network model with 5 configurations, the parameter reduction range is 74%-93%, the computation volume reduction range is 49.4%-89.6%, and the Map floats -3.6%-0.6% compared with before pruning. It is proved that the pruning scheme proposed in this paper has both large compression ratio, high precision and flexibility, and has important reference significance for the real-time target detection task on board.