Face Shape Classification For Face Recognition In Large-scale Database

Face recognition has become a hot spot of pattern recognition research in recent years. It can be applied widely in many fields including the management system for household registration, national defense, security and so on. However, as the scale of face database increases, face recognition speed and rate have decreased remarkably. Therefore, it has become an issue to increase recognition speed and rate in large-scale database.Face shape is a stable feature in face and its classification method is very simple. Based on this, a pre-classification based face recognition method is presented in this paper. According to the face shape, large-scale database is divided into several smaller sub-databases. On one hand, it will decrease the amount of data in face recognition so as to speed up the recognition. On the other hand, it will get rid of the potential non-candidates by using the characteristic of face shape so as to decrease the false acceptance rate.An anthropometry based method is adopted in face shape classification. First the face is detected and positioned. Next feature points are extracted and facial indexes are computed according to the feature points. Finally the face images are classified based on the facial indexes of them.Here are the details of several parts mentioned above:1)In the part of face detection, we compare all existing algorithms and choose Adaboost as our method. Results show that it has a high detection rate.2)In the part of feature points positioning, AAM is adopted to extract feature points in face images. We make landmarks on the training image and train them. The final result is good.3)In the part of face shape classification, since the face shape is gradually changed, there is no sharp boundary between each face shape. For this reason, we divide it into several classes with overlapped parts. Results show that it does decrease the false rate of face shape classification though more space is taken up.The overall experiment shows that the presented method can remarkably improve the recognition speed and rate of a face recognition system with large-scale database.