Graph Reconstructed Tensor Subspace Analysis

Tensor Subspace analysis (TSA) is typical tensor space analysis method, which can obtain the projection matrixes by building an adjacency graph to model the local geometrical structure of the manifold, and has been successfully applied in many practical problems. But the essential neighbor graph constructed suffers from the following issues: the k-neighbor is computed by the whole image matrix, which leads to the lost of the correlative columns information, the graph can not express the geometrical and discriminative structures of the original data space accurately. Based on a general platform of graph embedding, we proposed a graph reconstructed tensor subspace analysis (GRTSA) which will better express the spatial structure information of the original image matrices, preserving the corresponding inter-column locally information. Real face recognition experiments show the superiority of our proposed GRTSA in comparison to TSA, also for corresponding supervised and unsupervised extensions.