Research On Deep Learning Reconstruction Algorithm For Compressive Image Sensing

With the development of the era and the progress of science and technology,people's demand for information is increasing rapidly.The acquisition,compression,transmission and storage of massive data that carry information are facing tremendous pressure.The traditional data processing method is based on the Nyquist theorem.In order to reconstruct the original signal with high quality,the Nyquist theorem requires that the sampling rate must be at least twice of the signal bandwidth.In recent years,the theory of compressive sensing(CS)broke the limitation of sampling rate by the Nyquist theorem and made it possible that the sampling rate is less than twice of signal bandwidth.It can simultaneously sample and compress signal,and greatly saves the storage space of the sampled signal.And compressive sensing reconstruction algorithm is the key point of compressive sensing theory,which plays an important role in promoting the further practical application of compressive sensing.In this dissertation,the methods of deep learning have been applied to the reconstruction algorithm of compressive sensing,and the main research results are summarized as follows:(1)A compressive sensing image reconstruction algorithm based on convolutional neural network,Compatibly Sampling Reconstruction Network(CSRNet),is proposed.The algorithm is used to reconstruct high quality images from CS measurements sampled by a random sampling matrix.The CSRNet is non-iterative algorithm,which has low computational complexity,fast reconstruction speed,and can achieve better quality of reconstruction image at extremely low sensing rate.The framework of CSRNet adopts cascade connection mode,which includes three modules:preliminary reconstruction module,deep reconstruction module and residual reconstruction module,in order to gradually improve the quality of reconstruct image.The experimental results show that compared with the contrast method,the reconstructed image of the proposed algorithm has obvious improvement in both objective and subjective quality.(2)A compressive sensing image adaptive sampling and reconstruction algorithm based on convolutional neural network,Adaptively Sampling Reconstruction Network(ASRNet),is proposed.The ASRNet includes novel compressive sensing sampling module and reconstruction module based on neural network,which realizes the matching of sampling module and reconstruction module.It not only improves the compressive sensing sampling efficiency,but also improves the efficiency of image reconstruction,and greatly improves the quality of reconstructed image.The experimental results show that the ASRNet can obtain high quality reconstructed images compared with the reconstruction algorithm based on traditional sampling method.(3)A compressive sensing full image reconstruction algorithm based on convolutional neural network,Compressive Sensing full image Reconstruction Network(CSReNet),is proposed.There are two modules,reconstruction module and removal module in the CSReNet.Different from other deep network-based algorithms,the proposed algorithm can not only recovery image from compressive sensing measurements,also remove the blocky artifacts of reconstructed image to improve quality.Experimental results show that our proposed network outperforms several compressive sensing reconstruction algorithms both in objective and subjective visual quality.