Research Of Image Restoration Neural Network Based On Multi-layer Convolutional Sparse Coding

Image restoration has always been a basic and classic problem in computer vision,aiming at reducing or eliminating image degradation.In recent years,sparse model methods and deep learning methods have been widely used in the research of image restoration.Among them,deep learning methods based on convolutional neural networks have achieved better recovery performance than sparse model methods,but the lack of its theoretical basis has always been the problem of improving the network structure and optimizing the model.The method of combining the sparse model with the neural network provides a new perspective to solve this problem.In particular,the multilayer convolutional sparse coding has established a close connection with the forward pass of deep convolutional neural networks through its tracking algorithm,so that the improvement of the network can follow the guidance of the algorithm.Therefore,based on the learning iterative soft thresholding algorithm(ML-LISTA)of multi-layer convolutional sparse coding,this paper combines it with the learning mechanism of neural network to study the image restoration problems of fluorescence microscope image denoising and image super-resolution reconstruction.The main contents are as follows:1.The theory of compressed sensing,convolutional neural network,multi-layer convolutional sparse coding and multi-layer learning iterative soft threshold algorithm(ML-LISTA)are described,and theoretical analysis of the combination of multi-layer convolutional sparse coding model and convolutional neural network is carried out.2.A fluorescence microscope image denoising network(MCSC-net)based on multilayer convolutional sparse coding is proposed.The ML-LISTA algorithm for obtaining the deepest sparse representation in the multi-layer convolutional sparse coding model is embedded into the convolutional neural network,and an end-to-end supervised neural network(MCSC-net)with jump connection is constructed for real fluorescence microscope image denoising.The network can be naturally interpreted as a parameter optimization denoising algorithm,each layer of the network strictly corresponds to each step of the unfolded ML-LISTA algorithm processing,and the two parameters(number of iterations,number of filters)in ML-LISTA determine the depth and width of the network,which brings theoretical guidance for the improvement of the network.At the same time,dilated convolution is introduced into the network training to extract multi-scale context information without additional parameters,so as to improve the network denoising performance.The comparison experiments between the proposed method and other existing methods on real fluorescence microscope data sets show that under all noise levels,the peak signal-to-noise ratio(PSNR)of this method is improved by about 1 ～ 3 d B compared with traditional denoising methods;when the noise level is large,its PSNR is improved by about 0.1 ～ 0.4 d B compared with the deep learning method(Dn CNN).In addition,in terms of visual performance,our method can not only reduce image artifacts,but also better restore image details and edge contours,showing a strong appeal in actual denoising applications.3.An image super-resolution reconstruction network(SRMCSC)based on multilayer convolutional sparse coding is proposed.In view of the high similarity between low resolution image and high resolution image,and the significant sparsity of residual image,the ML-LISTA algorithm based on multi-layer convolutional sparse coding is combined with CNN while introducing residual learning.The ML-LISTA algorithm that solves the deepest sparse representation in the multi-layer model is used to extract the residual features,and then combined with the input image to reconstruct the high-resolution image.This not only speeds up the training speed and convergence speed,but also compares with a variety of existing excellent methods,the proposed method has a significant average PSNR value gain of about 0.2 ～ 1.1 d B under all scale factors.Especially compared with the deep learning method(SRCNN),when the data set is BSD100 and the scale factor is 2,the PSNR value improves about 0.7d B.At the same time,in terms of visual performance,this method can better restore the detailed texture of the image,which verifies the effectiveness of the proposed method in super-resolution reconstruction.4.The main work of this paper is summarized,analyzes and prospects the follow-up research of fluorescence microscope image denoising and super-resolution reconstruction under the combination of multi-layer convolutional sparse coding and convolutional network.