Improved median linear discriminant analysis for face recognition
Feilong Zhang, Xiaolin Chen, Bei Zhang, Shunfang Wang · 2013
Traditional linear discriminant analysis (LDA) exaggerates the contribution of distant samples in center calculation for identification, resulting in suboptimal shortcoming. This paper proposes an improved method based on LDA, which is named as KDA method in this paper because it gives different weights to different training samples according to K nearest neighbor idea in within-class scatter matrix calculation, and chooses K nearest classes among all to calculate the total center in between-class scatter matrix calculation. Considering the interference of outliers when sample size is small with high dimensional data, a new median discriminant algorithm (MDA) method is also proposed, which uses an improved median (not real median) to substitue the mean in center determination. Finally MDA and KDA are combined to form a MKDA method. The comparison among LDA, KDA, the new MDA and MKDA methods with ORL face database is given. Experimental results suggest MKDA performs best among the four and both KDA and MDA outperform LDA.