Image Compressed Sensing Using Non-Local Neural Network

Deep network-based image Compressed Sensing (CS) has attracted much attention\nin recent years. However, the existing deep network-based CS schemes either\nreconstruct the target image in a block-by-block manner that leads to serious\nblock artifacts or train the deep network as a black box that brings about\nlimited insights of image prior knowledge. In this paper, a novel image CS\nframework using non-local neural network (NL-CSNet) is proposed, which utilizes\nthe non-local self-similarity priors with deep network to improve the\nreconstruction quality. In the proposed NL-CSNet, two non-local subnetworks are\nconstructed for utilizing the non-local self-similarity priors in the\nmeasurement domain and the multi-scale feature domain respectively.\nSpecifically, in the subnetwork of measurement domain, the long-distance\ndependencies between the measurements of different image blocks are established\nfor better initial reconstruction. Analogically, in the subnetwork of\nmulti-scale feature domain, the affinities between the dense feature\nrepresentations are explored in the multi-scale space for deep reconstruction.\nFurthermore, a novel loss function is developed to enhance the coupling between\nthe non-local representations, which also enables an end-to-end training of\nNL-CSNet. Extensive experiments manifest that NL-CSNet outperforms existing\nstate-of-the-art CS methods, while maintaining fast computational speed.\n