Multi-Scale Residual Reconstruction Neural Network with Non-Local Constraint

Abstract

With the development of the neural network, some novel reconstructed networks are proposed to solve the problem of compressive sensing (CS) reconstruction. Compare with the traditional reconstruction algorithms, they can reconstruct the original images from the compressed measurement quickly and accurately with a low sampling rate. However, the CS reconstruction algorithms based on neural network ignore the image non-local similarity that is important prior information for the reconstruction. We propose a multi-scale residual reconstruction neural network with non-local constraint (NL-MRN). First, it considers the prior image non-local similarity and adds a non-local operation into the reconstruction network. Then, different scale residual reconstruction modules that have different convolution kernel size are combined to obtain the final output. Finally, the loss function of the whole network is defined as a weighted sum of the loss function of different scale reconstruction modules. What is more, the training efficiency of the network is improved by the proposed segmental training method. The theoretical analysis and the experimental results show that the proposed NL-MRN achieve better reconstruction compared with other reconstruction algorithms, especially at a low sampling rate.