Facial trait code and its application to face recognition

Abstract

We propose the Facial Trait Code (FTC) to encode human facial images. The proposed FTC is motivated by the discovery of the basic types of local facial features, called facial trait bases, which can be extracted from a large number of faces. In addition, the fusion of these facial trait bases can accurately capture the appearance of a face. Extraction of the facial trait bases involves clustering and boosting approaches, leading to the best discrimination of the human faces. The extracted facial trait bases are symbolized and make up the n-ary facial trait codes. A given face can be then encoded at the patches specified by the traits to render an n-ary facial trait code with each symbol in its codeword corresponding to the closest trait base. We applied FTC to a typical face identification problem, and it yielded satisfactory results under different illumination conditions.