Research On Restoration Of Non-uniform Motion Blurry Image Based On Strong Edge Extraction Network

Motion blur is a basic type of image blur.Motion blurry image is usually divided into uniform motion blurry image and non-uniform motion blurry image.Uniform motion blurry image refers to that each pixel of the blurry image has the same blur type and degree,that is,it has the same blur kernel.The non-uniform motion blurry image refers to the uncertainty of the blur types and degree of each pixel of the blur image,that is,the blur kernel type or sizes maybe different.Blind restoration of blurry image refers to the restoration of a clear image from a blurred-image when the blur kernel is unknown.A lot of work have been done on blind restoration of uniform blurry image,many methods of blur kernel estimation and image restoration are proposed.For non-uniform blurry image deblurring,because of the complexity of the blur kernel,the existing methods have great limitations and most even fail.In recent years,deep learning has been widely used in the field of image processing,including the restoration of blurry images.By elaborately designing the convolutional neural network,deep learining can recover the blurry images by end-to-end convolutional neural network without estimate blur kernels,it brings new ideas for the restoration of non-uniform motion blurry images.Although some achievements have been made in the deep-learning-based non-uniform motion blurry image restoration,their robustness is still poor.The reason is that the strong edge of blurry image is not restored effectively.So,this paper focus on the strong edge extraction method of blurry image for non-uniform motion blurry image deblurring.The proposed network is an end-to-end generation antagonism network.Therefore,the proposed method does not need to estimate the blur kernel of non-uniform blur image,which can greatly speed up the restoration of non-uniform blur image,and achieve a better restoration effect.The main work of this paper is as follows:(1)Research the status of blurry image restoration native and foreign.This paper analyzes the blur kernel characteristics of non-uniform motion blurry image and the basic methods of deep learning,deeply studies the methods about blind restoration of non-uniform motion blurry image based on convolutional neural network,and summaries the difficulties in research.(2)A blind restoration network of non-uniform motion blurry image based on edge constraint is proposed.The network is designed based on the generation antagonism network.The core part is generator network,which consists of edge extraction sub-network and blur feature extraction sub-network.The edge extraction sub-network is mainly used to extract the edge features of gradient domain from the blurry image.Because the gradient domain feature contains the edge information of blurry image,it is helpful to help and constrain the blur feature extraction sub-network and enhance the restoration effect of the blurry image edges.(3)A blind restoration method of non-uniform motion blurry image based on strong edge extraction network is proposed.The non-uniform motion blur restoration network based on edge constraint proposed in(2)still has many deficiencies,the edge feature extraction method is not perfect,the designed network is simple,and the training process is unreasonable.In order to solve above problems,a blind restoration method of non-uniform motion blurry image based on strong edge extraction network is proposed.This method is also an end-to-end blurry image restoration network based on generating antagonistic network.The generator network is composed of strong edge extraction sub-network,blur feature extraction sub-network and feature fusion sub-network.The strong edge feature extraction sub-network is used to extract the strong edge feature of blurry image and filter out the feature information irrelevant to the strong edge in the ordinary gradient image.The blur feature extraction sub-network uses the cross-resnet block disigned in this paper to enlarge the network's field and extract the blur features of the blurry image.The feature fusion sub-network mix the extracted strong edge feature map with the corresponding network layer feature map,constrains the parameter optimization process of the network,and with the help of the powerful feature automatic learning ability of the convolutional neural network,strengthens the network's adaptive restore ability to different types and degrees of non-uniform motion blurry images.(4)Designing and carrying out the relevant verification experiments.It is divided into two parts: one is to verify the effectiveness of the proposed method;the other is to compare with the current excellent non-uniform motion blurry image blind restoration method of traditional and based on deep learning.Experiments show that the proposed blind restoration method based on strong edge extraction is fast,high quality and robust.