Abstract: The resources of edge devices are limited, but the deep neural network model is large and has many parameters. Aimed at the difficulty of deploying the SAR image ship target detection model based on deep neural network on edge devices, a compression scheme based on pruning is proposed to lightweight the YOLOv3-SPP model. Experiments were performed on AIR-SARSHIP-1.0 data set using this method. Experimental results show that the original model can be compressed to less than 10% by the proposed compression scheme, while the speed is increased by 1.5 times and the accuracy is only slightly decreased by 0.02, which greatly reduces the difficulty of deploying the model at the edge and breaks through the limitations of hardware resources.