Image Compressed Sensing Deep Reconstruction Based On Multi-resolution Information Fusion

Different from Nyquist sampling theorem,which states that the sampling rate must be greater than or equal to 2 times the highest frequency of the signal,compressed sensing theory points out that if a signal is sparse in a certain domain of transformation,the signal can be accurately reconstructed from a small number of sampling values.As a new signal sampling method,compressed sensing provides a new criterion for signal and information processing,which is of crucial importance in reducing the cost of signal's storage and transmission,and shows great application prospect in the fields of communication,imaging and image processing.Compressed sensing reconstruction aims to restore the original measured signal from the reduced dimension sampling value,which is the core problem of compressed sensing theory.In recent years,the method based on deep network has shown good potential in compressed sensing reconstruction,and has obvious advantages over traditional model optimization methods.However,some existing deep reconstruction networks still have some shortcomings,such as too deep to train,poor quality of reconstructed image at low sampling rate,too much network computation and parameters.To solve the above problems,based on the analysis of the basic principle of compressed sensing,this paper focuses on the research of multi-resolution complementary reconstruction,discusses the parallel reconstruction of multi-resolution image,and uses relatively accurate low resolution image to guide the high-quality reconstruction of high-resolution image;at the same time,to improve the quality of the reconstructed images,discusses the cross-fusion of depth features of different resolution images to increase the integrity of information.The main contributions of this paper are as follows:1)Based on the discovery that multi-channel networks can better improve the feature representation capabilities of deep networks,a two-channel information complementary and enhanced compressed sensing deep reconstruction network is designed.Practically speaking,the reconstruction network consists of two parallel channels.One channel realizes the reconstruction of the original resolution,and the other channel realizes the reconstruction of the image by downsampling;the middle part uses the channel attention of the depth features of the low resolution image to assist the discrimination and selection of the high resolution channel features;finally,the two channels are fused to realize the complementary enhancement of the information.A large number of experiments based on the publicly released standard test set verify the effectiveness of this method in improving the image reconstruction quality of deep networks.2)Based on the discovery that the cross-fusion of different scales feature can effectively enhance the integrity of the information and thus improve the quality of the reconstructed image,a new method of image compressed sensing reconstruction with multi-resolution depth feature complementary enhancement is further proposed on the work of 1).Specifically,the deep reconstruction network we designed contains multiple channels to reconstruct the original resolution image and the down-sampled resolution image respectively;at the same time,the attention information extracted from the low-resolution channel is used to assist the feature selection of the high resolution channel,and the strategy of cross-fusion of the depth features extracted from different channels is used to enhance the feature representation ability of the deep network.A large number of experimental results verify the effectiveness of this method,and show that it can obtain better reconstructed image quality compared with the existing state-of-art methods.