Image Compressed Sensing Using Non-local Neural Network

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