Research On Hybrid Noise Removal Algorithm Based On Image Self-similarity And Convolutional Neural Network

In recent years,due to the development of computer vision,image processing technology has been used in various fields such as aviation,meteorology,medical treatment,and security.However,because the image is easily interfered by various noises during the process of acquisition,transmission and storage,the image quality will be reduced,which will directly affect the subsequent processing of the image.Therefore,image denoising is an essential part of image processing.At this stage,researchers are studying Gaussian noise and have proposed many high-performance denoising algorithms.However,the noise distribution of mixed noise is more complicated.Researches of mixed Gaussian and impulse noise are not deep enough,and the denoising performance of related algorithms needs to be improved.At present,most algorithms for mixed noise removal are based on a two-step strategy.That is,the first step is to detect and replace impulse noise,and then remove Gaussian noise.After the impulse noise is replaced,the distribution of mixed noise is still far from the Gaussian distribution.At the same time,the removal of mixed noise based on the two-step method largely depends on the detection accuracy of impulse noise.When the impulse noise is strong,the denoising performance will be greatly reduced.Aiming at the problem that the noise distribution is far from the Gaussian distribution,this paper proposes a non-local mean-based mixed noise removal framework—NMF.For the dependence of impulse noise detection,a deep neural network denoising algorithm based on hybrid attention mechanism—HACNN is proposed.The main contents of this dissertation are as follows:(1)Based on the two-step method,a non-local means based mixed noise removal framework—NMF is proposed.The impulse noise pixel are detected by median filtering,and replaced by its non-local means,so that the distribution of mixed noise approximately obeys the Gaussian distribution.Combining NMF with low-rank approximation and convolutional neural network,NMF-LR algorithm and NMF-CNN algorithm are given respectively to verify the effectiveness of NMF.In the NMF-LR algorithm,non-locally similar image blocks are combined into a matrix,and a low-rank approximation algorithm is used to reconstruct the denoised image.At the same time,by introducing the gradient regular term,the goal of better preservation of image details is achieved.In NMF-CNN,batch normalization and residual learning are introduced to improve the efficiency of training and the performance of denoising.Experiments show that compared with traditional mixed noise denoising algorithms,NMF-LR can achieve better denoising effects.At the same time,compared with the mixed noise denoising algorithm based on convolutional neural network,the NMF-CNN algorithm also has better denoising performance.(2)Based on the one-step method,a deep convolutional neural network algorithm HACNN based on hybrid attention mechanism is proposed for mixed noise removal.Compared with the existing two-step denoising algorithm,HACNN removes the mixed noise of multiple noises in only one step.In HACNN,on the one hand,batch normalization is used to improve the training efficiency of the denoising model,and on the other hand,hybrid attention is introduced to achieve the purpose of better retaining image texture details.Experiments show that for mixed Gaussian and salt and pepper noise,the experimental results of HACNN can approximate the experimental results of the best existing algorithms.For more complex noises(such as Gaussian mixture,salt and pepper and random value impulse noise,etc.),the HACNN algorithm proposed in this paper not only has better denoising performance,but also has better visual effects on image details and textures than all other competing algorithms.