Design and Implementation of Neural Network Accelerator Based on SAR Image Target Detection Algorithm

In this paper, a SAR image target detection acceleration technology based on FPGA hardware resources is proposed. Using RTL code level design method, four basic convolutional neural network operators are developed: conv2d, PW, DW and reshape. At the same time, using software and hardware collaborative design method, a set of reconfigurable and reconfigurable convolutional neural network accelerator based on AI instruction set is designed to support mobilenetv3 Yolov3 target detection network and realize the task of low-power and high-precision target detection for SAR images. The measured results show that the neural network accelerator can reach 102 FPS with mobilenetv3 network structure and 125 FPS with yolov3 network structure.