Research On Blind Image Restoration Based On Convolutional Neural Network

Image blind deblurring refers to recovering a clear original image from the observed degraded image when the point spread function is unknown or part of the information is known.It is a solution to the ill-posed inverse problem.In the computer vision field,image blind deblurring is a challenging problem.The fuzzy categories are roughly divided into three categories:motion blur,defocus blur,and Gaussian blur.Mainly for the image Gaussian blur problem,this article uses the deep learning method to retrieval the blurred image.The main contents are as follows:Firstly,a method of recovering Gaussian blurred images using u-net multi-scale network structure is proposed,which avoids the process of alternately estimating blur kernels and images in some traditional algorithms,and reduces the complexity of calcluation.The network adopts the end-to-end training method,which solves the problem that there will be a large number of‘artifacts'in the image recovery due to incorrect blur kernel estimation.And we compare different loss functions that effect the ability of the network.The experiment results show that the u-net multi-scale network structure can better preserve the image detail information,and the₁l-norm loss function is better for image restoration.Secondly,the edge information of the image,which is remained by the traditional algorithm,is beneficial to the estimation of the image and blur kernel.Two sub-networks are added to the multi-scale network u-net.The blur image is used to learn clear image gradients through the sub-network,and then the gradient information is introduced as a prior into the u-net multi-scale convolutional neural network.Experiments show that after the gradient prior is added,the image restoration quality is better than the result without gradient prior.When using different subnets,residual blocks is superior to the common convolutional layer.Finally,in order to further improve the quality of image restoration,a conditional adversarial network framework was proposed.The adversarial network consists of a generative network and a discriminate network,which also called a generator and a discriminator.The generator is the network structure mentioned in Chapter 3,and the discriminator is a network consists of convolutional layers and activation functions.The experiments results show that using adversarial network helps to improve the image reconstruction quality and achieve better results.