Deep Convolutional Neural Network For Removal Of Salt And Pepper Noise In Image

Image denoising is a common problem in image processing.The first step in many image processing processes is to denoise the obtained image.Salt and pepper noise is a special case of image noise.It randomly converts certain pixel values of the original image to 0 or 255 to contaminate the image.The traditional image denoising algorithm is based on filter design or interpolation algorithm.However,as far as we know,there is few research on using end-to-end convolutional neural networks to remove salt and pepper noise.This dissertation mainly uses the convolutional neural network model in deep learning to remove salt and pepper noise.First,we review the application of traditional denoising algorithm and deep learning denoising algorithm.The dissertation focus on two deep learning models: VGG and Dn CNN.Secondly,we designed different strategies and trained different denoisers according to different image situations.Then we selected the appropriate denoiser.Finally,we conducted a large number of experiments on different datasets for different images,and demonstrated the robustness of our denoiser.In addition,because the deep neural network has a wider local receptive field,it has a better advantage in the removal of large density noise.The main work is as follows:1.The related applications of image denoising are reviewed,including traditional methods and deep learning methods.The evaluation criteria of image performance in image denoising are given.The related theory of convolutional neural network,VGG model,Dn CNN model and Matconvnet framework are introduced.2.In order to be able to select appropriate denoising methods for different levels of noise,the dissertation trains different denoisers based on deep convolutional neural network.First,in the data collection phase,the public dataset is used for training.During the training process,the image is rotated and flipped to achieve data enhancement.In the model verification phase,12 images are used as verification dataset to prevent over-fitting for selecting the appropriate denoiser.In the model test phase,we test the classic images and images from different datasets.The experimental results show that the choice of denoiser has a good performance in denoising.3.During the experiment,our choice of denoisers did not achieve a good denoising effect for images with a large number of interference points.In order to further investigate the denoising ability of the denoiser,we selected more datasets.We designed an algorithm to judge the number of image interference points and applied different denoisers separately for images with more interference points.