Face Recognition Research Based On Decision Fusion And Distance Metric Learning

The study on Automatic Face Recognition (AFR) has both significant research value and bright prospect of application. Since1960s, in the filed of constrained AFR, great progresses have been made continuously and many practical applications have also emerged. However, in the filed of unconstrained AFR, popular algorithms which are effective under constrained condition cannot achieve the same performance. Therefore, how to design robust face recognition algorithm under unconstrained condition is a new challenge in computer vision world.This work is dedicated to make some contributions to solve this problem. The main contributions of this work are listed as follows:(1) A decision fusion based face recognition algorithm framework is proposed. Due to the great variations of face pose, illumination, occlusion and face expression in unconstrained scenario, it is impractical to utilize only one kind of face feature in face recognition algorithm. According to this notion, this work uses several complementary face features in order to make a complete representation of human face. At the decision level, this work proposes a method to select several sub-classifies which have small combinational correlation to form a strong classifier. The final strong classifier displays better performance than any sub-classifier.(2) Distance metric learning method is used in this work to learn Mahalanobis distance to replace the ordinary Euclidean distance. After utilizing distance metric learning method, the distance between the face images of the same identity is decreased, while distance between face images of different identities enlarged. Therefore, metric learning method is helpful to improve the performance of classification.(3) To further enhance the performance, a "divide-and-conquer "method to explicitly cope with the large variations of pose is also proposed. In this work, a pose estimator is applied to put all the training and testing samples into different pose categories. In each category’, classifier training and testing are processed independently. As a result, every classifier is only trying to handle one kind of pose combination. In this way, the accuracy of classifiers is improved.The performance of the algorithms proposed in this work is tested under LFW regulation. The results show that the decision fusion is better than feature fusion. Furthermore, distance metric learning and pose estimation both promote the performance of face recognition algorithm. Therefore, in the specific scenario of face recognition, distance which is learned from face samples should be used and face images with large pose variation need different features to represent them.