Research Of Face Image Super-resolution Method Based On Representation Learning

Face super-resolution technology refers to the low-resolution face images through technical processing to get high-resolution face images,which has been widely used in many aspects,such as transmission of face images,artificial intelligence,face image processing in criminal investigation cases and so on.Due to the fact that a large part of the practical application is to synthesize highresolution one for a single low-resolution face image,compared with the traditional multi-frame reconstruction based face super-resolution method,now the learning-based face super-resolution method,which makes use of the priori information of face images to synthesize a high-resolution face Image,is the research hotspot.Therefore,the main works of this dissertation are as follows:Firstly,a new algorithm based on kernel locality-constrained adaptive iterative neighbor embedding(KLAINE)is proposed.The algorithm takes into account that in the process of iterative process of face image synthesis,the best neighbor value of each time is changed,so the strategy of adaptively selecting neighbors in high-resolution training samples is proposed.Then,the lowresolution face image to be synthesized and the k-nearest neighbor low-resolution training samples are mapped into the same high-dimensional space to achieve the synthesis of non-linear features.Finally,the entire low-resolution training samples which are updated by KLAINE algorithm are utilized to synthesize a new high-resolution face images again.The final high-resolution face image can be obtained by several updating progresses.Secondly,a locality-constrained iterative matrix regression based on nuclear norm(NLCIMR)is proposed.Since all the previous methods convert the face image matrix into a vector,the intrinsic structure of the image block is destroyed,so the proposed algorithm directly uses the image block matrix to participate in the synthesis.In addition,the nuclear norm constraint reconstruction error can more accurately describe the distribution of the error matrix and the introduction of iterative ideation in the synthesis process is to further improve the robustness of the algorithm.Lastly,a multi-pose algorithm based on double nuclear norms regularization(DNRMP)is proposed.In order to solve the problem that face images acquired in practical application environment often have a variety of pose,the algorithm corrects the face images using orthogonal Proclock first and enhances the robustness of face images synthesis with pose changes.The first two algorithms proposed above are validated in the FEI face database and the real environment face database.The last one is verified in the FERET face database.Compared with the contrast algorithm,these algorithms show some advantages and effectiveness.