Cross-Scale Dilated Residual Network for Image Compressed Sensing

Abstract

Deep Learning based Compressed Sensing (DCS) algorithms are able to accomplish better rebuilding images compared to classical Compressed Sensing (CS). However, the majority of DCS algorithms focus on recovery and information transfer over network depth, while ignoring the communication between networks and losing some information about the traits. To enhance the exchange of data between networks of different scales, the cross-scale dilated residual network (CDRNet) is proposed. In the reconstruction section we use a parallel network and incorporate a dilated residual block (DRB) as a method to expand the convolutional field to obtain features at different scales, and then a cross-scale information exchange block(CIEB) to superimpose the information at different scales to achieve a better detailed reconstruction performance. We experimentally compare the CDRNet $*$ (without CIEB) and the CDRNet (with CIEB). The results show that information exchange helps image reconstruction, and our algorithm achieves better quality of rebuild and texture recovery than other CS algorithms.