Fusing local appearance models for face recognition

Face recognition is a popular research area due to its scientific challenges and potential applications. Therefore, it attracts attention from both diverse research communities and the industry for several years. Various techniques have been intensively investigated to provide a robust face recognition system. Many of these techniques have already achieved very high recognition accuracies under controlled conditions. On the other hand, face recognition under uncontrolled conditions is still a very hard problem. The difficulty arises from facial appearance variations caused by various factors, such as expression, illumination and partial occlusion, and the time gap between training and testing data capture. Face recognition based on local features usually outperforms holistic approaches because local representation is less sensitive to appearance variations caused by occlusions and facial expressions. In this thesis, a local appearance based face recognition algorithm, which works reliably under real-world conditions, is proposed. The proposed algorithm uses different local appearance models to represent face images. Fundamentally, it exploits Gabor features that have been extensively used for facial image analysis due to their powerful representation capabilities. It utilizes curvature Gabor features in addition to conventional Gabor features. The system focuses on selecting and combining multiple Gabor classifiers. The final Gabor classifier is obtained by LLR-based fusion of classifiers that are selected using SFFS-based classifier selection algorithm. In addition, the system uses DCT features as extra evidence. Finally, classifiers trained on different local representations are combined at score-level by PLSR-based fusion. The system is evaluated on FRGC version 2.0 Experiment 4, and achieves 94.16% verification rate @ 0.1% FAR, which is the highest accuracy reported on this experiment so far in the literature.