Hierarchical face recognition using an adaptive discriminant space

In this paper, we propose a novel method using an adaptive discriminant space to identify people's faces from various directions. The adaptive discriminant space that is constructed from the candidates' facial images optimally distinguishes between faces among certain candidate.. A principle method of our research gradually limits the classification of categories by using face direction estimation and recognition. Both the face direction estimation method and the face recognition method are appearance-based methods that employ linear discriminant analysis (LDA) on the four-directional features (FDF). First, our method chooses the candidates using a hierarchical combination of face direction estimation and recognition. Next, face direction is estimated exactly by each candidate's discriminant space of face direction estimation. Finally, our method creates a new discriminant space from the candidates' facial images selected from the results of the face direction estimation procedure. Limiting variations of a face direction can strengthen a face recognition discriminant space Using the adaptive discriminant space, the candidates can be selected optimally. Experiments showed that our method improved our accuracy rate 1.2 % to 98.8 %, achieved by hierarchical face recognition of 105, 000 images from 150 subjects facing 35 different directions.