Research On Motion Blur Restoration Network Based On Multi-Stage

In the process of acquiring images on a daily basis,the influence of external factors such as object movement or camera movement leads to image distortion in the captured photos.Motion blur is a common image degradation problem,and motion blur image recovery techniques have been a hot research topic in the direction of image processing.Researchers classify image deblurring methods into two categories according to whether the blur kernel is known or not: blind recovery methods for blurred images and non-blurred recovery methods for blurred images.In this thesis,we deeply investigate the blind recovery technique of motion blurred images,analyze the defects of existing methods in detail,and propose a new multi-stage based motion blur recovery network to improve the recovery quality of reconstructed images.The advanced features of the proposed network are mainly reflected in the following aspects:(1)A multi-stage motion blur recovery network model is constructed to address the detail loss problem of traditional deblurring methods.In this thesis,considering the detail loss problem of traditional deblurring methods and referring to the multi-stage detection algorithm of target detection algorithm,the deblurring task is innovatively decomposed into multiple stages,which are feature extraction stage,artifact removal stage and detail recovery stage.The detail recovery stage uses the reconstructed image from the remove image artifacts stage as input and performs secondary reconstruction by adaptively fusing the discovered high-latitude features through simultaneous convolution to obtain the final reconstructed image.(2)Numerous frontier network models are studied and their own network models,optimization strategies and loss functions are constructed.Firstly,in this thesis,an optimized pre-trained residual network is used in the feature extraction stage,a multi-scale network with weight sharing and cross-stage feature fusion is used in the remove image artifacts stage,and an adaptive synchronous convolutional detail recovery network based on the idea of attention is used in the detail recovery stage.Secondly,the nonlinear activation-free block with semi-normalization function is used in the proposed network in this thesis to improve the learning ability and generalization of the network.Finally,this thesis adopts Adam W optimizer for gradient optimization in the network,and combines Charbonnier loss function,edge loss function and Fourier loss function as the overall loss function of the network,so that the gradient direction of network convergence is better and the training effect of the network is improved.(3)The evaluation criteria of conventional reconstructed images are investigated,and the recovery quality of reconstructed images is evaluated using the image pixel histogram combined with two conventional evaluation criteria,namely peak signal-to-noise ratio and structural similarity,to better evaluate the performance of the network.This thesis demonstrates the rationality of the proposed network through several experiments.The experiments and data analysis provide theoretical support for the proposed network,which is validated and tested on several datasets.The proposed network is validated and tested on several datasets.The proposed network has implications for research directions such as deep learning and image restoration.