Tensor Discriminant Color Space for Face Recognition

Recent research efforts reveal that color may provide useful information for face recognition. For different visual tasks, the choice of a color space is generally different. How can a color space be sought for the specific face recognition problem? To address this problem, this paper represents a color image as a third-order tensor and presents the tensor discriminant color space (TDCS) model. The model can keep the underlying spatial structure of color images. With the definition of n-mode between-class scatter matrices and within-class scatter matrices, TDCS constructs an iterative procedure to obtain one color space transformation matrix and two discriminant projection matrices by maximizing the ratio of these two scatter matrices. The experiments are conducted on two color face databases, AR and Georgia Tech face databases, and the results show that both the performance and the efficiency of the proposed method are better than those of the state-of-the-art color image discriminant model, which involve one color space transformation matrix and one discriminant projection matrix, specifically in a complicated face database with various pose variations.

I. Introduction

As one of the hottest topics in the field of pattern recognition and artificial intelligence, face recognition has been widely used in public securities, such as crime and terrorist detection, etc. There are various subspace transformation methods for recognizing faces. Principal component analysis (PCA) [1] is a widely used linear subspace transformation method maximizing the variance of the transformed features in the projective subspace. Linear discriminant analysis (LDA) [2] encodes discriminant information by maximizing the between-class covariance, while minimizing the within-class covariance in the projective subspace. Moreover, in order to keep spatial structure information of a gray image, Yang et al. [3] proposed an algorithm called two-dimensional PCA (2D-PCA) for face recognition, in which the image covariance (scatter) matrix is directly computed from the image matrix representation. Li and Yuan [4] extended this idea using discriminant information and presented 2D-LDA, which constructs the image between-class covariance matrix and the image within-class covariance matrix. All these methods are used to deal with gray face images rather than color face images because some past researches suggested that color appears to confer no significant face recognition advantages beyond the gray [5].