Natural Scene Image Denoising Based On Deep Learning

With the advent of the information age and the popularization of various digital products,people often use digital images as a carrier to communicate with each other,and image processing technologies can be seen everywhere.The image is inevitably disturbed by various noise sources in the process of acquisition,transmission and preservation.In order to get the original information more accurately,the image denoising is the key.Image denoising,as a basic task in digital image processing,provides a solid foundation for digital image analysis and understanding.In the field of image processing,many scholars have proposed many excellent image denoising methods,which can effectively denoise images,such as median filtering and Gaussian filtering.However,these traditional image denoising methods mainly face a specific type of noise,which assumes that the noise in the image is a single salt-and-pepper noise or Gaussian noise.And noisy image SNR can not be too low,once applied to the natural scene,the image denoising effect dropped sharply,almost impossible to meet the application needs in practice.Therefore,looking for a common,fast and effective image denoising algorithm is a problem worth studying.This paper analyzes and researches the related techniques of image denoising,the effect of traditional denoising methods and the latest research progress.Based on the shortcomings and deficiencies of the prior art,a set of image denoising methods based on depth learning and image noise classification module is proposed in natural scenes.The main work is as follows:(1)The related knowledge of digital image denoising and the common denoising filtering algorithms are introduced.The advantages and disadvantages of the mathematical characteristics of noise distribution,the classification of noise and the traditional filtering algorithms are summarized.(2)An image noise classification module based on residual network is proposed,and the feature vector of image noise is obtained by using the ability of CNN to obtain the image abstract features from the bottom.In order to improve the accuracy of the noise classification,this paper uses a network structure with a residual module,so that the entire image noise classification module can work on a deeper network structure.Experiments show that the module has a good ability to distinguish the phenomenon of image noise.(3)An image denoising module based on self-encoder network and anti-neural network is proposed.Among them,for the Autoencoder Net,a structure with a jump connection is proposed,which greatly improves the convergence speed of the Autoencoder Net.At the same time,the loss function of WGAN network is improved to meet the demand of image de-noising task,which can better preserve the image details.(4)Aiming at the difficulty of training in deep neural networks,the problem of over fitting is prone to occur,and training accelerating strategies and overfitting strategies are provided to facilitate the faster training of neural networks.Experiments show that the image denoising method used in this paper can effectively deal with a variety of imagenoise pollution,while for mixed noise images,the effect can be more obvious than the traditional method.