Research On Motion Blurry Image Restoration Methods Driven By Mathematical Model And Data

With the widespread application of imaging devices such as digital cameras and smartphones,high-quality and low-cost digital images have become essential tool for exchanging information,which are widely used in security monitoring,medical assistance,crime tracking,and other applications.However,during image capturing,relative motion and depth of field changed between the captured object and the imaging devices are unavoidable,which would easily cause motion-blurry problem.Since image quality is essential for analyzing and understanding the subsequent image data,the problem of motion-blurry image restoration that can restore the corresponding clear image from the motion-blurry image has received extensive research attentions.However,most of the motion-blurry image restoration methods have some problems,such as bad image-prior description,inaccurate extraction of significant edges,and poor interpretability of network-structure design.Therefore,according to different scenes,how to perform image-prior knowledge portrayal and design network model,and to propose effective mathematical model-driven and data-driven image restoration methods are important topics that require in-depth research and have a wide range of application values.To solve the above issues,this paper studies mathematical model-driven and data-driven blurry image restoration methods.Based on prior knowledge description,significant edge extraction,semantic information expansion and model guide,efficient motion-blurry image restoration methods are proposed.Furthermore,the image restoration network solves the restoration of blurry images with low resolution,and an interpretable model for restoring blurry image is designed.The main content of this paper is divided into the following aspects:1.Aiming at the problem of an insufficient image prior knowledge characterization in the motion-blurry image restoration,a motion-blurry image restoration framework based on priorknowledge description is proposed.A blurry image restoration method based on Elastic-net prior and significant edge structure is proposed.The Elastic-net prior is applied to the motionblurry kernel estimation method.The significant edge structure of the image is used to improve the estimated accuracy of the blurry kernel.Furthermore,considering the influence of significant noise about image restoration,a fractional dark channel prior is given to obtain effective natural image constraints.The gradient prior is used to accurately estimate the image blurry kernel,which can achieve motion-blurry image restoration and improve restoration results' visualization effect.2.Since the salient edge extraction will affect the accuracy of blurry kernel estimation,this paper proposes a motion-blurry image restoration algorithm based on the edge structure extracted by the encode-decode network.The training and learning method is used to construct a codec network for the extracting salient edge structure,After considering various regularization constraints,the motion-blurry image can be restored by combining the salient edge structure.The proposed method avoids complex threshold selection for edge extraction and reduces the computational complexity of the algorithm.3.Aiming at the problem of poor applicability of edge information about low-resolution blurry images,a low-resolution blurry image restoration method based on semantic information is proposed.Based on the data-driven framework,the semantic information is extracted from the low-resolution blurry image.After that,the semantic information is used to modulate the deep neural network's feature information which restores the high-resoluition clear image from the low-resolution blurry face images.At the same time,it ensures the consistency of the face image's identity information before and after restoration.4.Due to the poor interpretability of the network design based on the black box principle,a model-guided design method of the motion-blurred image restoration network is proposed,which is applied to the motion-blurry face image restoration.We firstly research the mathematical model-driven method for motion-blurred image restoration.Further,the general approximation theory of convolutional neural networks is used to study the equivalent representation method between the neural network structure and the traditional mathematical model.And then,a model-guided network is built to restore the motion-blurry images.Experimental results show that the motion-blurry face image restoration network based on model guide can avoid potential model parameter redundancy,improve model calculation efficiency,and have an excellent restoration effect on motion-blurred face images.