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

Image denoising has always been considered as a research hotspot in the field of image processing,how to recover the potential clean image from the contaminated image has been widely concerned by many scholars.Among many denoising methods,some methods based on convolutional neural networks have gradually become the mainstream with their strong learning ability,but the lack of complete mathematical explanation and theoretical guidance is inevitable in the process of improving such denoising methods.In recent years,the multi-layer convolutional sparse coding model(MLCSC)has builded close connection with the forward pass of convolutional neural network through pursuit algorithms(such as multi-layer learned iterative soft threshold algorithm,ML-LISTA),which makes it possible to design and understand convolutional denoising network from the perspective of algorithm.In this paper,based on ML-LISTA algorithm to obtain the deepest sparse representation vector of MLCSC model,combined with network op-timization technologies and weighted recursive supervision mechanism,three recursive denoising convolutional neural networks with interpretable structure and clear mathe-matical significance are proposed for image denoising.The main contents of this thesis are as follows:1.In the adaptive learning framework of convolutional neural network,a novel re-cursive denoising convolutional neural network RDn CN-LISTA is proposed by fusing the sparsity of image representation coefficients.It obtains the deepest sparse representation vector of the MLCSC model through forward pass,updates the convolutional dictionaries automatically through back propagation,and finally completes image denoising through sparse representation,which successfully integrates sparse prior into the denoising con-volutional network.The end-to-end network structure is composed of embedding net,inference net and reconstruction net,which makes every layer and every skip connection of this network strictly correspond to the update steps of ML-LISTA algorithm.It is more meaningful than designing network structures through experience and experiment.The depth and width of the network are determined by two network super parameters:the number of recursion and the number of filters,which are chosen as 11 and 128 by experiment.Then,the denoising experiments on Set12 dataset show that the basic net-work can achieve the same denoising results as classical denoising network Dn CNN when the additive white Gaussian noise(AWGN)level is 50,and the average PSNR value is27.23 d B.2.To further enhance the denoising performance of the network,combined with the network optimization technologies,dilated convolution and batch normalization are intro-duced into the first three layers of the embedded sub through experiments,the optimized recursive denoising convolution neural network was proposed,dubbed as RDn CN-LISTA~+.In the real image denoising experiment,compared with other classical denoising methods,our network benefits from the sparse prior of image itself,which makes it more robust to noise interference.The best denoising results are obtained when the level of AWGN is 25,50,and75,which verifies the denoising performance of this network.In particular,the average PSNR values of Set12 and BSD68 are 0.31db and 0.18db higher than that of Dn CNN.Although the denoising performance is general when the noise level is 15,which reveals the limitations of deep recursive network,the image”house”still obtains high-quality image with PSNR value of 35.15d B.3.To avoid the limitation caused by the incomplete utilization of deep feature information and the potential gradient explosion(disappearance)of recursive network,a weighted recursive denoising convolutional network WRDn CN-LISTA~+with recursive supervision mechanism is proposed.By moving the reconstructed net layer to the end of each recursion,the result of each recursion is the estimated denoised image.The results of each recursion are transmitted to the last layer of this network by many skip connections with some learnable weights.All recursion results are fully used to improve the quality of the final denoised image,and effectively supervised.The experimental results show that:the introduction of recursive supervision mechanism weakens the impact of different recursive times,and improves the deficiency of the RDn CN-LISTA~+model when the noise level is low.The network got the best denoising result at all noise levels,especially when=15,which is better than the classical Dn CNN.The average PSNR values of Set12 and BSD68 are improved by 0.03d B and 0.05d B.From the perspective of subjective observation,this method can better maintain the details of the image texture,reduce artifacts and distortions.It is an effective denoising network.