A compression method for object detection network for edge devices

In order to solve the problem of insufficient computing power of edge computing equipment, this paper draws on the lottery ticket hypothesis theory, through unstructured global pruning of the network and using the pruning network of the previous stage to sparsely distill the pruning network of the current stage, while maintaining the network sparsity. At the same time, the heuristic information of the pruning network of the previous stage is used to help the pruning network recover the parameters that were wrongly pruned, and then the two are combined to obtain the output network of the current stage, so that the unstructured pruning network is obtained after several rounds of iteration. Then it is pruned by structured filter to enable the network to combine the advantages of structured pruning and unstructured pruning, so as to take into account the accuracy and acceleration effect of the network. It is verified on the yolov7 algorithm. The accuracy is almost unchanged compared with the method before pruning, and the speed is improved by more than 45 %, which has achieved very good performance.