Dynamic Weight of Adaptive Information Density Network for Image Super-Resolution

Abstract

A model algorithm based on image information density characteristics is proposed to achieve network structure adjustment. The image dense region classification is fed back to the subsequent network. According to the classification information, the image sampling window is sent to different network to realize pixel-level channel switching, thereby reducing the network deployment process's computational pressure. The dynamic weighting network adjusts the weight coefficients of pixels in the sampling window to approximate the image's shape and generate better texture effects than FSRCNN. When using the public test sets to evaluate the adaptive information density network structure, the computation complexity of SRCNN and FSRCNN was reduced by about 28%, and the PSNR only reduced by about 0.1dB.