Research On Blind Image Restoration Based On Multi-scale Convolutional Neural Network

Blind image restoration,is recovery of potentially sharp images from known quality-degraded images,where the point spread function is unknown.It is one of the important research directions for image deblurring.At present,the traditional image blind restoration algorithm requires a complicated iterative calculation process,and the restored image quality is low.In order to improve the performance of image blind restoration method,this paper applies multi-scale convolutional neural network to study this problem.The specific research contents are as follows:Firstly,this paper proposes a multi-scale image deblurring method based on dilated convolution to expand the receptive field.This method designs a small convolution module with different parameters to increase the receptive field and extract feature information of different scales.The mapping between blurred images and sharp images is learned by training large amounts of data.Experimental results show that the results obtained by this method have higher image clarity.Secondly,in order to further improve performance,this paper proposes a multi-scale image deblurring method based on total variation(Total Variation,TV)loss.Inception-ResNet(Residual Inception Net,Inception-ResNet)module and Residual network are used by this method to construct the network,the gradient information is used to calculate the loss function.The network input has three dimensions,so both the loss function and the network structure use a multi-scale structure to better convey image information.Experimental results show that the method can significantly improve the image quality.Finally,in order to obtain more realistic details,this paper proposes the image deblurring method based on adversarial generative network.Three generator network structures are designed using Inception-ResNet module,attention mechanism module,upsampling and downsampling.The three methods are evaluated on a large-scale deblurred public data set with complex motion.The experimental results show that the proposed method has better performance in both visual and peak signal-to-noise ratio metrics,and has better visual effects on deblurring of actual blurred images.