Research On Single Image Deblurring Method Based On Deep Learning

With the continuous popularization of intelligent devices,people frequently contact images in their daily lives.Images contain a large number of information elements,and humans can communicate information through images as a carrier.In daily life,when we use a mobile phone or camera to take photos,the resulting image is not a real-time image in a single time,but a scene over a period of time.During the process of taking photos,the camera will generate exposure.In these short seconds,the displacement of the object being taken or the displacement of the camera itself may cause pixel confusion,resulting in a non-uniform blur,Causing image information to be damaged and unable to transmit effective information.Deblurring has always been a relatively difficult problem in the field of image processing.This topic focuses on this kind of ambiguity,analyzes the shortcomings of current non-uniform blind deblurring algorithms in dynamic scenes,and proposes a large coding and small decoding network structure based on multi-scale residual attention blocks.The structure design of a single image deblurring network is implemented through a rough-to-thin tactics.The algorithm improvement of this topic mainly covers three aspects:(1)By studying the problem of image detail loss caused by insufficient feature extraction capability in the coding part of the network,a large coding structure is constructed,and a multi-scale residual block integrated with attention mechanism is designed to heighten feature extraction and heighten the repair of rich details.In each coding block,three consecutive multi-scale residual blocks integrated with attention mechanism are used.First,a 5 × 5’s convolution layer to draw the initial features,and then input these features into the multi-scale residual attention block,employ diverse convolution kernels to draw multi-scale features,and combine two attention mechanisms to achieve more adaptive deblurring effect under different color channels and spatial distribution,achieve correction of feature space and different color channel areas,and better improve image detail recovery,Figures out the question of image detail loss attributed to weak feature extraction ability.(2)By studying the problems of small receptive field and incomplete global information in the network,resulting in inconsistent texture information and unclear pattern of the restored image,a module for expanding receptive field is constructed.A parallel expansion convolution with multiple expansion rates is designed in the coding block,and the receptive field is amplifyed without increasing the size of the kernel and features are extracted from objects of different scales,Let convolution export include a large scope of message,capture multi-scale fuzzy object information,improve the global information,extract patterns containing rich texture information,improve the quality of image restoration,and solve the problem that the texture of the restored image is inconsistent due to the small field of perception,and is visually unsightly.(3)By studying the problem that the low quality of the transmitted features in the network leads to artifacts in the restored image,a block for improving the effectiveness of the feature transmission is built between each two stages,and a feature filter block is designed.The local features are reweighted by using the on-site monitoring mechanism,and the extracted features are filtered to achieve the suppression of features containing less information,improve the effectiveness of the input features,and reduce artifacts,Achieve better deblurring effect.Solves the problem of artifacts in the restored image due to low quality or invalid feature transmission.By comparing the lately arithmetic with the arithmetic in Go Pro dataset,Real Blur-R in low-light situation and Kohler dataset,the effectiveness of the arithmetic proposed in this paper is proved,and the algorithm proposed in this paper has a relatively good performance in repairing image details and removing artifacts.