The Research Of Image Denoising Algorithm Based On Low-Rank And Deep Learning

As an important carrier of information storage and transmission,images are also an important way for humans to perceive and identify information.However,due to various factors,images are often contaminated by noise.In order to restore the real information of the image as much as possible,scholars have proposed a lot of denoising algorithm such as traditional denoising algorithm,non-local denoising algorithm and deep learning denoising algorithm.Although there are a large number of denoising methods have been applied,the research of this issue remains to be investigated.In this dissertation,the non-local denoising algorithm based on low-rank clustering and the denoising algorithm based on deep learning are studied.The non-local denoising algorithm is used in small data and deep learning algorithm is used in sufficient sample.The main work and innovation of this dissertation are as follows:(1)The classical image denoising algorithm is analyzed,including traditional spatial domain and transform domain algorithms,non-local self-similarity algorithms such as NLM,BM3 D,NCSR,and deep learning algorithms such as CSF and TNRD.Combining with the objective evaluation index PSNR and the subjective evaluation method of image quality,the performance of the denoising algorithms are assessed through experiments.(2)Based on the low rank theory,the WNNM(Weighted Nuclear Norm Minimization)algorithm of low rank matrix approximation is studied and the following improvements are proposed.a)For the problem that the image block structure information is not considered by using the Euclidean distance metric image block similarity in similar block aggregation process,the image structure information based on wavelet pre-filtering and structural similarity is introduced.That is a new block distance metric is proposed in order to increase the accuracy of the similar block aggregation matrix;b)For the denoising image noise residual phenomenon,the image is iteratively denoised by back projection technology to enhance the performance.Finally,the performance of the improved algorithm is evaluated by comparison with other denoising algorithms,and the effectiveness of the improved method is verified.(3)Based on the deep learning technique,the DnCNN denoising model of convolutional neural networks is studied.Furthermore,an image patches optimized model based on multi-flow and multi-scale dilated convolution is proposed.The improvements are as follows: a)Introducing multi-scale dilated convolution structure to make full use of image features to enhance the performance of the network model;b)The current neural network denoising model is only suitable for the problem of certain Gaussian noise.Therefore,a multi-flow multi-scale network structure based on image patches is proposed to realize blind noise denoising.The effectiveness of the improved method is verified by comparing the performance of it with other classical deep learning denoising algorithm.At the same time,in the experiment,above method is applied to the mixed noise images with non-unique variance,and predominant results are obtained.