Lightweight Image Compression Sensing Network With Group-based Super-resolution Reconstruction

As a new sampling theory,compressed sensing points out that when the signal itself or in a certain transform domain is sparse,the signal can be compressed and sampled simultaneously,and can be reconstructed from the sample values that the number is far lower than that of Nyquist sampling theorem required.Compressed sensing is of great significance to reduce the cost of signal transmission and storage,and has broad application prospects in optical imaging,medical imaging,image processing and communication.Compressed sensing includes two processes: sampling and reconstruction.Reconstruction is the core issue of compressed sensing theory.Compared with traditional methods,deep learning methods that have been emerging in recent years have obvious advantages in improving the quality and speed of compressed sensing reconstruction.In this paper,the low-complexity implementation of compressed sensing reconstruction is studied to better meet the needs of many mobile applications.Compressive sensing reconstruction methods based on deep learning mostly include initial reconstruction and deep reconstruction.Most of the initial reconstruction is realized by single matrix multiplication operation or single-layer convolutional network with high complexity,which leads to huge resource consumption and high computational complexity.Therefore,this paper first discusses how to effectively reduce the complexity of initial reconstruction by combining grouped reconstruction with low-resolution reconstruction.Based on this,deep super-resolution reconstruction is further explored to better reduce the complexity of reconstruction network and maintain the quality of reconstruction.The main works of this paper include:(1)A lightweight image compressive sensing reconstruction method based on grouping initial reconstruction is proposed.This method aims to greatly reduce the complexity of the initial reconstruction module through the combination of grouping reconstruction and low-resolution reconstruction.Specifically,under the framework of joint optimization sampling and reconstruction,the input sampling values are divided into multiple groups to recover multiple low-resolution initial estimates of the original images in parallel.Then,an up-sampling technique is used to map the multiple low-resolution estimates to an initial reconstruction image with the same resolution as the original image.Finally,deep features are extracted by deep network and the details of reconstructed image are estimated.Experimental results indicate that the proposed method can effectually reduce the complexity of the initial reconstruction network while maintaining or even enhancing the quality of the reconstructed image.(2)A deep compressive sensing reconstruction method based on super-resolution reconstruction is proposed.The basic idea of this method is to extract the deep features of the image directly in the low-dimensional image space,and then get the high-dimensional features and high-resolution reconstruction result by up-sampling,so as to further reduce the complexity of the reconstruction network.Specifically,on the basis of work(1),the deep enhancement reconstruction network is designed to firstly extract low-dimensional deep features of images directly from the multiple low-resolution initial estimates,then obtain high-dimensional deep features of images through up-sampling,and finally obtain reconstructed residual images through a simple transformation.At the same time,the residual branch with gradient feature fusion is introduced into the deep enhancement reconstruction network,which can enhance the learning and representation ability of the network by fusing gradient features in different directions.Since the introduction of gradient operator does not bring additional parameters,the reconstruction performance can be effectively improved without increasing the parameter size.Theoretical analysis and a large number of experimental results show that the proposed method can greatly reduce the complexity of the system while ensuring the performance of reconstruction.