Research On Image Compression Sensing Algorithm Based On Deep Learning

Compressed sensing(CS)theory has created a new way of signal observation and has injected new vitality into the field of signal processing.In recent years,the emergence of deep learning theory has brought new ideas for compressed sensing recovery methods and promoted the development of deep compressed sensing research direction.However,the current widely used image depth perception network models have two limitations: first,most deep learning methods are single-scale models,which lead to the limited reconstruction performance of images;second,most deep learning-based methods only utilize the observations in the initial reconstruction and do not fully utilize the information carried by the observations.This paper aims to solve these problems and selects a topic to study compressed perception methods based on deep learning,and the main work accomplished is as follows:(1)To address the problem that single-scale deep learning models limit the image reconstruction performance,a multi-scale compressed depth perception algorithm composed of lightweight spatial attention modules(LWSA)is proposed.First,the whole network consists of multiple cascades of lightweight spatial attention modules,where the attention modules consist of cascades of inflated convolutions with different expansion coefficients for better reconstruction using multi-scale features;second,to eliminate the tessellation lattice effect caused by the inflated convolutions,the output of each inflated convolution takes fusion processing and focuses each module feature on the key spatial content through feature clustering.Finally,the proposed network is validated on three datasets,set11,set5,and set14,and the results show that when the observation rate is 0.1,the improvement over the reference CSNet and CSNet+ is 0.83 dB and 0.54 dB on dataset set11,0.72 dB and 0.65 dB on dataset set5,respectively,and 0.75 dB and 0.62 dB on dataset set14,respectively,and the visualization results also show that the extracted network achieves good results.(2)A parallel enhancement network of observation residuals(RP-CSNet)is proposed to address the problem that most compressed-aware deep learning methods only use observations in the initial reconstructed image without fully utilizing the information properties of the observations,resulting in a poorly reconstructed image.First,the network uses two parallel branches to extract features,one of which uses a larger convolution kernel to propose the main features,while the other uses observation residuals to compensate for the missing local features.Second,feature fusion is used at the front,middle and end of the network to achieve full depth fusion of multi-feature data.To improve the network performance,interpolation compensation is applied in the process of main feature and local feature extraction,and the features extracted from the two branches are compensated for each other separately.Experiments show that the proposed observation residual compensation shows better performance at observation rates of 0.01,0.04,and 0.1,and improves about 0.2-0.6 dB in PSNR over the reference method on data sets set11,set5,and set14,which illustrates the effectiveness of the network.