Abstract:

This paper proposes a binary structure feature selection (BFS) for face recognition. In the proposed method, all classes are combined in pairs. Based on the two-class classifier, the most suitable features for discriminating these two classes are chosen to form a feature-selected space. During the test on an unknown sample image, similarities between the unknown image and all training classes are calculated in the featureselected space. The unknown image is thus judged to belong to the class which shows the highest similarity. Performance of the method has been tested with the AT&T and AR face databases. The results show that, compared with other methods, the proposed technique can achieve higher recognition rate with a low feature dimension.

Key words:  face recognition, principal component analysis, linear discriminant analysis, feature extraction