Face Compressed Image Restoration Method Based On Subspace Regression Learning

The restoration of face compressed images is essentially a non-deterministic problem of solving linear equations.Some prior knowledge of face compressed images can be used to cluster the image patches into different subspaces,so as to narrow the solution range as a constraint.The main research results of the face compressed image restoration method based on subspace regression learning proposed in this thesis are as follows:1.Face compressed image restoration method based on structure-dependent subspace regression.Aiming at the problem that the pose of face images is not fixed in the real scene,a face compressed image restoration method not limited to the face pose is proposed,which mainly includes training and restoration.In the training phase,using the facial landmark localization,the compresseduncompressed image patch pairs can be divided into subspaces related to the face structure,and the projection matrix of the compressed image patches to its uncompressed version can be learned in each subspace.In the restoration phase,for each compressed input image patch,the structuredependent subspace to which it belongs is determined,and an appropriate linear mapping of this subspace is selected to obtain an estimated restored version.The experimental results show that the proposed method is superior to other comparison algorithms in both subjective and objective evaluation.2.Face compressed image restoration method based on hierarchical subspace regression.In order to improve the accuracy of subspace clustering,a face compressed image restoration method based on hierarchical subspace regression is proposed,which mainly includes training and restoration.In the training phase,the rule of the face edge-orientation distribution is used to classify the compressed-uncompressed patch pairs into shallow subspaces.Then,the k-means clustering is used to cluster the deep subspaces of each shallow subspace,and corresponding mapping matrix training is performed for each deep subspace.In restoration phase,an appropriate linear mapping kernel selected based on the edge-orientation of compressed input image patch is applied to generate the restored output image patch.The experimental results show that the proposed method is superior to other comparison algorithms in both subjective and objective evaluation.