Uncorrelated Discriminant Vectors vs. Orthogonal Discriminant Vectors in Appearance-Based Face Recognition
Fengxi Song, Rubin Zheng · 2009
Uncorrelated linear discriminant analysis (U-LDA) which seeks a set of statistically uncorrelated discriminant vectors has been pursued by many researchers in the field of face recognition. Unfortunately, it has two inborn deficiencies. First, it provides little new knowledge in addition to conventional LDA since discriminant vectors of Fisher linear discriminant are usually statistically uncorrelated. Second, it is based on an unreliable intuition that removal of statistical correlation between discriminant vectors is favorable for pattern recognition. From experimental studies conducted in the paper we found that U-LDA methods could be significantly inferior to their orthogonal counterparts in face recognition. Our work implies that U-LDA methods might be futureless in face recognition.