Research on Compression Optimization Algorithm for Super-resolution Reconstruction Network

Abstract

Under the condition of limited resources of embedded systems, the paper proposes a compression optimization algorithm based on pruning and quantization, so that the computational requirements of the super-resolution reconstruction algorithm based on a Convolutional Neural Network (CNN) can be met. First, a multiple regularization pruning optimization algorithm based on an attention module and a BatchNorm layer is proposed. Then, a coordination optimization algorithm of INT8 training and quantization for FPGA architecture is proposed. The performance of the pruning optimization algorithm was verified for the Super-Resolution CNN (SRCNN), the Fast Super-Resolution CNN (FSRCNN), and the Very Deep Super-resolution CNN (VDSRCNN). As for SRCNN, the performance of the quantization optimization algorithm was verified on the FPGA EC2 hardware simulation platform. The results show that the proposed compression optimization algorithm can achieve a good balance between network accuracy and inference speed.