Two-dimensional uncorrelated linear discriminant analysis for facial expression recognition

Abstract

The uncorrelated discriminant linear analysis (ULDA) has been proved to be an effective feature extraction method and is known as a development of classical linear discriminant analysis (LDA). In real-world applications, we often encounter the "small sample size" (SSS) problem that the number of training samples is less than the dimension of feature vectors. Under this situation, the within-class scatter matrix is always singular, making the direct implementation of the ULDA algorithm inapplicable. To tackle this problem, it is common to apply a preprocessing step that transforms the data to a lower dimensional space with loss of valuable information contains in original data. In this paper, a new technique called two-dimensional uncorrelated linear discriminant analysis (2D-ULDA) is developed for solving the SSS problem. The main ingredient is the small size of covariance matrix which is suitable for the SSS problem. To evaluate the performance of the proposed 2D-ULDA, a series of experiments were performed on JAFFE database. The recognition accuracy across all experiments was higher using 2D-ULDA than ULDA. The comparison experiments of the proposed 2D-ULDA, 2DPCA and 2DFLD also demonstrated the competitiveness of our approach.