Research And Application Of Motion Blurred Image Restoration Algorithm

Image restoration refers to restoring low-quality images degraded by various factors into high-quality images that meet the requirements of various applications.The blurred image caused by motion has a serious impact on subsequent high-level vision tasks such as object detection and object tracking.In recent years,important progress has been made in the research of motion blurred image restoration based on deep learning.But in the real world,it is very difficult to obtain blurry and sharp image pairs with exactly the same lighting,composition,color and other factors.Therefore,based on machine learning theory,the study of unsupervised motion blur image restoration method has important theoretical significance and application value.Based on machine learning theory,this paper studies motion blurred image restoration algorithms for blurred image sequences and unpaired blurred clear images,respectively.Firstly,starting from the motion cues of blurred image sequences,the spatial variation blurring and edge restoration are studied.Then,starting from unpaired blurred clear images,the role of structural information in the restoration of motion blurred images in the Cycle GAN model is studied.Finally,an application system integrating motion blur image restoration algorithm and restoration image quality evaluation method is designed and developed.The main work of this paper includes:(1)A self-supervised motion blur removal algorithm based on edge and multi-scale features is proposed.For spatially varying blur,obtaining a larger receptive field and multi-scale information is crucial for improving the quality of restored images.Based on this,this paper introduces a atrous residual module to obtain a larger receptive field and a context module to fuse multi-scale features,and designs an edge loss function to improve the model’s attention to edge information.Experiments on the Fastec＿synthetic＿dataset show that the performance of the self-supervised motion blur removal method proposed in this paper is close to that of supervised methods and outperforms other comparative unsupervised methods.(2)An unpaired image motion blur removal algorithm based on structural information is proposed.When the Cycle GAN model is applied to the motion blur image restoration task,the generator tends to learn the color,texture and other information of the clear image to "trick" the discriminator,while the removal of motion blur information is not enough.Aiming at the above problems,this paper proposes an unpaired image motion blur removal algorithm based on structural information.The edge map of the generated image is sent to the discriminator together with the generated image to increase the network’s attention to the removal of motion blur information.Compared with the base network,the improved model improves the PSNR index by1.37 d B and 2.29 d B on the text dataset and the face dataset,respectively,and the SSIM index by 0.0885 and 0.0484 respectively.The PSNR index is improved by 3.57 d B on the high-speed railway tip rail telescopic displacement dataset.(3)According to the application requirements of image restoration,a motion blur image restoration system is designed and developed.The above motion blur image restoration algorithm is integrated into the system,and an image clarity evaluation module based on edge information is designed to quantitatively and objectively evaluate the image restoration results of different algorithms,and push the results to the user in order.This paper contains 34 figures,15 tables,and 60 references.