Directionally classified eigenblocks for localized feature analysis in face recognition

Abstract

A new local feature extraction method is introduced. The directionality of local facial regions is regarded as essential information for discriminating faces in our approach, which is motivated by the directional selectivity of the Gabor wavelet transformation, which has been preferred to others for face recognition. The discriminative directional information is forced to be compacted in a few coefficients by applying principle-component analysis with the support of directional classification in the discrete cosine transform domain. The local features extracted by our method are better at discriminating face patterns than previous ones, as was verified by comparison of class-separability results. Also, in face recognition simulations using rigid and flexible face matching strategies based on locally extracted features, our proposed method showed outstanding performance. (c) 2006 Society of Photo-Optical Instrumentation Engineers.