Face Recognition Using Direct-Weighted LDA

This paper introduces a direct-weighted LDA (DW-LDA) approach to face recognition, which can effectively deal with the two problems encountered in LDA-based face recognition approaches: 1) Fisher criterion is nonoptimal with respect to classification rate, and 2) the "small sample size" problem. In particular, the DW-LDA approach can also improve the classification rate of one or several appointed classes by using a suitable weighted scheme. The proposed approach first lower the dimensionality of the original input space by discarding the null space of the between-class scatter matrix containing no significant discriminatory information. After reconstructing the between- and within-class scatter matrices in the dimension reduced subspace by using weighted schemes, a modified Fisher criterion is obtained by replacing the within-class scatter matrix in the traditional Fisher criterion with the total-class scatter matrix. LDA using the modified criterion is then implemented to find lower-dimensional features with significant discrimination power. Experiments on ORL and Yale face databases show that the proposed approach is an efficient approach to face recognition.