Spaceborne SAR Target Recognition Method Based on Edge Computing

Aiming at the problems of the traditional satellite image processing system, such as the extension of the image pro-cessing task and high pressure of broadband data transmission between the satellite and the ground, this paper adopts the idea of edge computing to extend the SAR image processing site from the ground data processing center to the mobile edge processor on the satellite. Taking full account of the limited resources on the satellite, this paper uses channel pruning and knowledge distillation to compress the convolutional classification network. Under the condition of ensuring little accuracy degradation, the all-convolutional networks (A-convnets) is compressed by 14.66 times and the VGG16-LRN is compressed by 109.43 times. ASIC inference chip for edge computing with lower power consumption is used for deployment. Compared with GPU deployment, ASIC chip deployment can obtain more than twice the inference speed with 0.5 times the load power consumption.