Research On Image Denoising Method Of Salt And Pepper Noise Based On Deep Convolution Neural Network

In the process of digital image acquisition and transmission,various image noises occurred due to the influence of external environment interference or equipment component failure.Salt and pepper noise is of a common type,which appears as black spots or white spots on the image.The visual quality of image and the application of image in engineering fields are seriously affected,so the removal method of images with salt and pepper noise has been widely studied and applied.At present,the removal methods of salt and pepper noise are mainly divided into two categories: one is the traditional filtering algorithms,the other is the deep learning based algorithms.Traditional filtering algorithms rely on the detection accuracy and the reconstruction accuracy of noise pixels.At present,such methods have great limitations in improving the denoising performance.The denoising method based on deep learning can obtain significant advantages in image denoising performance by training an end-to-end deep convolution neural network.In this thesis,a salt and pepper noise denoising network based on noise attention is proposed to remove salt and pepper noise in gray images and color images,which further improves the denoising performance based on the existing similar algorithms.The main research work is summarized as follows.(1)Inspired by the typical deep denoising network Dn CNN,a salt and pepper noise denoising network model based on noise attention is proposed.The improvement of network model is mainly in two aspects.Firstly,mark the noisy pixels and clean pixels,and combine the marking map and noise map as the network inputs.Due to the one-to-one correspondence between the marked map and the noisy points,this method makes the network perceive the difference between the noisy points and the noise-free points in the fitting training process,so as to pay more attention to the special noisy pixels,which is conducive to improving the recovery accuracy to the noisy pixels.Secondly,two symmetrical jump connections are set in the middle of the network.In the process of back propagation of network training,this jump connection makes the previous layers of the network easier to optimize,and more image information can be transmitted to the back layers.The experimental results show that the denoising performance of this salt and pepper noise denoising network based on noise attention is better than other related existing algorithms,and the restoration effects of image structure and texture details are also better than other related existing algorithms.It has obvious advantages in three denoising performance indexes: peak signal-to-noise ratio,structure similarity and feature similarity.Aiming at the problem that the gray denoising model in removing color image noise is not ideal,the denoising methods of color image are studied in this thesis.The existing denoising methods for color image are the denoising algorithms of gray image or the deep learning based model training on color image.The results of this thesis show that if the gray denoising model is directly used to denoise the three color channels of R,G and B,the correlation between the color channels will be ignored and good denoising effects can not be achieved.Therefore,this thesis adjusts the network input and output structure.Adjust the input channels of the gray-scale denoising network to 6,corresponding to R,G and B channels and its’ noise marking channels respectively,and adjust the output channels to 3.The experimental results show that the color denoising network can significantly improve the denoising performance of color images after training the color image datasets.(2)Aiming at the problem that the deep denoising model is sensitive to the noise density level and has insufficient generalization ability,a salt-and-pepper noise removal method for a wide range of noise levels is studied in this thesis.The existing network structure and training methods have strong pertinence for image noise density.In order to improve the generalization ability of the denoising network,a shallow noise density estimation network is designed firstly,and two solutions for a wide range of noise levels are proposed secondly.The first method is to add the noise density matrix corresponding to the noise image at the input of the network,and the noise image,the noise density matrix and the noise label matrix are combined as the network inputs;the second method is to detect the image noise density first,and then according to the density value to selecs the corresponding denoiser.The experimental results show that for an unknown noise density,both denoising methods for a wide range of noise levels can achieve good denoising performance,and the denoising performance of method two has a slight advantage over method one.