Research On Image Compressive Sensing Reconstruction Algorithm Based On Structure Sparse Deep Network

With the successful application of convolutional neural network technology in image classification,face recognition and other visual fields,it is of great significance to combine traditional related theories with deep learning technology in solving the problem of compressed sensing image reconstruction.In the deep learning-based compressed sensing reconstruction problem,due to insufficient feature extraction in the image,the reconstructed image is smooth and the high-frequency edge texture information is lost.Secondly,at a lower sampling rate,the sampled signal contains less information,resulting in information loss and blurring of the reconstructed image.If the prior information of various structures in the image can be effectively utilized,it will be of great help to the image reconstruction.Under the block-based compressed sensing framework,this paper combines deep learning technology with prior information of structural sparsity in images to guide the design of deep network structures and related loss functions.At a low sampling rate,the constraints on the reconstructed image solution space are realized,the high-frequency information of the image is effectively reconstructed,and the image reconstruction accuracy is improved.The research contents are as follows:(1)Combining image block structure sparse prior,image non-local self-similar prior and deep network,an image compressive sensing reconstruction algorithm based on structure group sparse network is proposed.First,use the observation similarity measure to construct similar groups for image observations,and input the similar groups into the two-branch convolutional neural network;input the observation similar groups into the edge contour reconstruction branch,and pass the local residual recurrent network and sub-pixels.The convolution layer obtains the reconstruction of the image edge contour;the similar groups of observations are input into the local detail reconstruction branch,and the reconstruction of the image detail texture is obtained through the densely connected network and multi-scale encoding and decoding network modules;finally,the two The branch reconstructed images are fused,and the output is the optimal reconstructed map of the original image.The structure group sparse constraint loss is designed and used to constrain the network training,which improves the stability of the network training and enhances the uniqueness of the image solution space.The experimental results show that,compared with the existing compressed sensing image reconstruction methods based on convolutional neural network,the proposed method can effectively reconstruct the edge contour and detail texture information of the image,and the reconstructed image has better visual effect.(2)According to the image non-local mean theory,and using the image non-local self-similar structure sparse prior to guide the deep network structure design,an image compressive sensing reconstruction method based on non-local feature fusion network is proposed.The method uses the complementary information between similar image blocks to synergistically represent single image features;a two-stage reconstruction network framework is designed with the idea of coarse to fine.In the first stage,a collaborative reconstruction group is constructed for the compressed observations of each image block,and a linear mapping network is used to obtain a coarse-precision initial reconstructed image group;A global residual reconstruction network formed by stacking adaptive interaction modules can accurately reconstruct the details of the image blocks to obtain the final output image.Finally,a patch sparsity constraint loss is constructed and employed to constrain the network training.The experimental results show that using the image non-local self-similar structure sparse prior collaborative image block for information reconstruction,the image evaluation indicators PSNR and SSIM have been greatly improved under the lower sampling rate of the reconstructed image.