Classical vector subspace learning algorithms work with vectorized representations of data manifold

while data manifold represented in the reality is intrinsically a tensor

so the algorithms based on tensor subspace learnig can perfectly detect the intrinsic geometrical structure of the data manifold.In this paper

a novel tensor subspace learning algorithm

tensor locality discriminant projection

is proposed.To implement the algorithm

construct within-class and between-class graph at first

then preserve local structure of the data manifold and utilize its discriminant information to deduce the formula of the algorithm

finally work out optimal tensor subspace by iteratively computing the generalized eigenvectors.The experiments on the standard face database demonstrate the effectiveness of the proposed algorithm.