Image super-resolution reconstruction algorithm based on channel shuffle

Abstract

Convolutional neural networks have achieved good results in single image super-resolution. Nowadays, image super-resolution model based on deep learning are gradually becoming lightweight, and lightweight networks use grouped convolution stacking, which hinders the channel Information flow between the two, and weaken the feature representation. Aiming at the existing problems of the existing image super-resolution lightweight model, this paper proposes a super-resolution reconstruction algorithm based on channel shuffle. This paper introduces a channel shuffle mechanism after grouped convolution. Channel shuffle disrupts the grouping order and allows group convolution to obtain input data from different groups, which allows the fusion of feature information between different channels. At the same time, this paper also introduces the dynamic activation function DY-RELV-B, which enhances the feature representation ability of the network model and further enhances the image reconstruction effect. Experimental results show that compared with LapSRN, VDSR, traditional interpolation method, etc., the method in this paper is superior to other methods in PSNR on x2 scale and x4 scale.