Joint Denoising And Super-Resolution Deep Network Via Generative Adversarial Training

Image denoising and image super-resolution reconstruction are setting on the core foundation of image processing technology.It has important applications in many fields such as defense and military,remote sensing navigation,and map mapping.However,facing low-resolution images with mixture noises,the traditional methods on either image denoising and image super-resolution alone cannot well handle the practical challenges.Image denoising is a process that reduces or eliminates noise in an image towards recovering high-quality image.The traditional denoising technique is mainly based on filter operations.Due to the complexity and diversity of the noise types,different filter-based processing methods can only handle specific kinds of image noise.Super-resolution reconstruction refers to the process of restoring low-resolution images to high-resolution images.Conventional super-resolution reconstruction is mainly based on various interpolation algorithms,which however have different effects on reconstruction of images in different scenes.Coming with the popularity of deep learning in image processing,it is of great significance to investigate image denoising and super-resolution reconstruction jointly using deep learning methods.Therefore,research on simultaneous image denoising and super-resolution retains as an open problem with significant challenges.In this dissertation,we study the problem of image denoising and super-resolution reconstruction.We contribute in the following three aspects:(1)For image with the mixed Gaussian white noise,salt and pepper noise and Poisson noise,this paper designs a deep neural network model to complete the mixed denoising task.In this model,the distribution of mixed noise is fitted based on the residual block structure of the residual network.The Inception structure is introduced into the residual block to reduce the number of parameters of the model.This model can effectively weaken the phenomenon of the disappearance of the model gradient,reduce the training difficulty and computational complexity,and remove the mixed noise well.(2)In the super-resolution reconstruction task,a neural network structure based on sub-pixels is designed to reduce the efficiency of the complex neural network structure.This structure achieves better super-resolution reconstruction effects with fewer parameters and high-scale factors.(3)We model the joint image denoising and super-resolution problems as a joint image restoration process.To this end,we use the generative adversarial network(GAN)to unify them into an "end-to-end" model.The model uses the generator to collaboratively perform image denoising and super-resolution reconstruction.The discriminator is used to discriminate the authenticity of the results,thereby simplifying the steps of image denoising and super-resolution reconstruction.Finally,comparative experiments show that the model has good performance and generalization ability.