Image Compression Sensing Reconstruction Based On Multiscale Variational Algorithms And Depth Convolution Neural Network

According to the compressive sensing theory,the original high-dimensional signal can be reconstructed from a small number of low dimensional observations with very high probability when the signal is compressible or it can be sparsely represented in a transform domain.This paper focuses on variational method-based,regularization model-based,and deep convolutional neural network-based compressive sensing method.The main work is summarized as follows:In this paper,the image is decomposed into structure component and texture component by using the stratified variational image decomposition algorithm.And then both of the structure and texture components are reconstructed by the hybrid basis-based CS algorithm.In order to enhance the sparsity of each component,we construct the TFWBST(Tight framework wavelet-based shearlet transform)further,which is combined with the wave atom to form a joint sparse dictionary.With the same sampling rate,our method can preserve more image details than WSGSR algorithm.Image prior knowledge is of great importance for solving inverse problems such as compressive sensing(CS).In order to regularize the image reconstructed from measurements,A shape-adaptive non-convex low-rank model is presented.To characterize the local property,the LPA-ICI(Anisotropic Local Polynomial Approximation-Intersection of Confidence Intervals)is used to analyze the local structure of each pixel and obtain its shape-adaptive neighbors.On the other hand,the non-convex low-rank model characterized by the l_p norm of singular values is used to formulate the non-local similarity.The combination of the shape-adaptive neighbors and non-convex low-rank model can characterize the local and nonlocal properties of natural images simultaneously,thus enforcing powerful constraints on the reconstructed images and producing competitive performance.Experiments on commonly used test images show that the presented method for image CS is effective with different sampling rates.A deep convolutional neural network(CNN)-based quality enhancement method is developed for images reconstructed from CS measurements.The presented quality enhancement CNN takes the recovered image produces by a traditional CS method as input and outputs a quality-enhanced image.Firstly,three groups of kernels with different sizes are used to extract features at different scales and these features are concatenated to generate a group of multi-scale features.Secondly,these multi-scale features are continuously transformed using a series of convolutional layers to produce multi-level features.Thirdly,low-,medium-and high-level features are combined and fed into a convolutional layer to predict a residual image.Finally,the estimated residual image is added to the network input to generate the final quality-enhanced result.The presented quality enhancement network not only uses advanced techniques including residual learning,batch normalization,and ReLU,also combines the multi-scale and multi-level features.Experiments show that the present method can improve the quality of the reconstructed images produced by traditional CS methods.The quality-enhanced images have fewer artifacts and more fine details.In addition,the presented quality enhancement method is efficient in testing phase.