The Research And Application Of Face Super-resolution By Deep Collaborative Representation

In the actual monitoring system,because the distance from the camera and the shooting device itself has a limited resolution,resulting in a small face image size that is difficult to identify.Therefore,how to improve the resolution of the captured face image,the quality of the face image,and improve the face image clarity and identification value become the urgent problem to be solved.Face Image super-resolution technology is one of the more economical and effective means to solve the above problems.Face image super-resolution is reconstructed the high-resolution(HR)face images from the input of single or multi-frame low-resolution(LR)face image.The technology is widely used in video surveillance,face recognition and even entertainment software.In recent years,benefiting from the rapid development of machine learning theory and application,the learning-based face super-resolution algorithm has become the mainstream direction of the research on the face super-resolution algorithm.The basic idea is to learn low-resolution images and high-resolution images mapping relations.This thesis focuses on the non-local similarity of human face images and the structural features of image blocks.The main results of research are as follows:(1)The face super-resolution by deep collaborative representation: the algorithm extends the traditional single-layer collaborative representation into deep model,it updates weight coefficient and corresponding HR and LR dictionary pair layer by layer,reconstruct deep linear function fitting complex HR and LR relation model.The algorithm is used to learn the dictionary pair in the feature domain,and the representation dictionary is updated layer by layer,and the manifold consistency of the HR and LR representation coefficients is improved,and the SR reconstruction performance of the face image is improved.Experimental results show that the proposed algorithm is superior to frontal face super-resolution algorithm in both subjective and objective evaluation performance in FEI face database.(2)Adaptive and weighted collaborative representation for face super-resolution algorithm: There are differences in the texture features for each region of the image.The proposed algorithm introduces an adaptive matrix and weight matrix to adaptively adjust the proportion of reconstructed error terms and expression coefficients in the optimization function which can reconstruct the target high-resolution images better.The experimental results show that the proposed algorithm outperforms the forerunner's face super-resolution algorithm both in subjective and objective evaluation performance on CAS-PEAL-R1 and LFW face database.In summary,this paper proposed two new face image super-resolution algorithm based on super resolution algorithm learning on the image,effectively overcomes the defects of traditional algorithm of the existing shortcomings and deficiencies,the high resolution face image clearly obtained and improves the quality of reconstructed images.