Ensemble Global And Local Features For Face Recognition

Face recognition has been one of the most active research topics in computer vision and pattern recognition. It has held more and more researchers'interest in recent years. Meanwhile, as a biometric technique, face recognition has wide range of potential applications in both public security and private life. Currently, the state-of-art face recognition system achieves satisfied results under the well-controlled environment. However, many evaluations and practices demonstrate that the performances of face recognition systems decrease tremendously under the uncontrolled imaging environment. Furthermore, in order to develop robust and practical face recognition system, many key problems needed to solved, especially the design of effective feature extractor and efficient classifier.Centering about the problem of facial feature extraction, this dissertation proposes to combine global and local features by a hierarchical manner, by which both the accuracy and the speed of face recognition system can be greatly improved. On the whole, the contributions of this dissertation are summarized as follows:(1) Propose a novel Gabor-based face recognition method. In recent years, Gabor feature has been recognized as one of the most effective face representation methods. However, the high dimensionality of Gabor features makes the following classifier design very difficult. Aiming at this problem, this dissertation proposes to group Gabor features into several feature vectors, then design a classifier for each feature vector. Finally, these classifiers are combined together to get the final classification result. The merit of this method lies in the"divide and conquer"strategy by which more discriminatory information is reserved and meanwhile the dimensionality of features fed into each classifier is greatly reduced. Experimental results show that this method is robust the variations of lighting, expression, partial occlusion and aging. Besides, compared with some popular face recognition methods, this method gives much better performance.(2) Propose a face recognition method by hierarchically combining global and local features. In the literature of psychology and neurophysiology, many studies have shown that, in human visual system, both global and local features are adopted for face recognition but they play different roles. Inspired by this phenomenon, this dissertation proposes to hierarchically combine global Fourier features and local Gabor features. Specifically, in the first layer, the global classifier is used for coarse classification. In the second layer, the global and local classifiers are combined together for more accurate classification. Experimental results show that the hierarchical combination of global and local features greatly enhances both accuracy and speed of face recognition system.(3) Propose a novel framework of linear feature extraction. Extracting discriminatory features is crucial for pattern classification tasks, but how to develop algorithms for effective feature extraction remains a challenging problem. Among numerous linear feature extraction methods, Fisher's Linear Discriminant Analysis (FLDA) is probably one of the most widely used approaches. However, FLDA has some inherent problems in both theory and practice. Based on the in-depth analysis on these problems, this dissertation proposes a novel framework for extracting linear discriminatory features, i.e., Optimal Discriminatory Projection Pursuit (ODPP). This framework consists of two steps: one is the generation of candidate projections (features); the other is the pursuit of optimal discriminatory projections. Experiments on both synthetic and real data sets show that the performance of ODPP is better than FLDA and its variants.In conclusion, through above-mentioned work, this dissertation makes an intensive study on the facial feature extraction. The experimental results show that: global and local features contain different discrmintory information for face recognition. By further extracting discriminatory features and hierarchically combining them, excellent system performance can be achieved.