Research On Face Super-resolution Method Based On Manifold Structure Preserving

Face super-resolution is a technology that reconstructs a high-resolution face image from a low-resolution one.It has important application value in areas such as surveillance and security.Since most of the applications in the actual situation are super-resolution processing for a single image,compare with the method based on multi-frame reconstruction,the learning-based method of reconstructing a single image using prior image knowledge is the current hot spot.(1)Considering the original high-resolution image blocks that are not affected by the degradation process,the manifold structure is more reliable,and the kernel function that can map features to high-dimensional space to achieve nonlinear feature synthesis.A face super-resolution algorithm based on a locality-constrained kernel bi-layer representation is presented.In this algorithm,the high-resolution layer is added to the model as a constraint term to compensate for the lost detail information.Then the high-and low-resolution features are mapped into the high-dimensional space,so that some nonlinear high-frequency features become linearly separable.The algorithm has been tested in the FEI database and the CAS-PEAL database.(2)A locality-constraint kernel couple-layer representation based contextual patch face hallucination method is proposed,which introduces the context relationship between image blocks.The algorithm adds high-resolution layer constraint and seeks similar blocks in high-resolution sample space to improve the problem of losing high-frequency detail information during degradation.Using kernel techniques to map feature information to high-dimensional space to obtain nonlinear characteristics.The algorithm has been tested in the FEI and CAS-PEAL database respectively for the no-noise and the noisy case to verify the effectiveness of the algorithm and its robustness to noise.(3)An adaptive selection based kernel context-patch face hallucination method is proposed.The contextual information is added to expand the receptive field of the image patch and enhance the global characteristic of the local-based method.The adaptive selection of nearest neighbor value K is used to make the selection of similar sample blocks of low-resolution input image block more reasonable,and avoids the problem of under-fitting or improper fitting.Finally,the kernel function is used to make the nonlinear features in the original space become linearly separable.The algorithm has verified in the FEI and CAS-PEAL database.