Biometric technologies have attracted much attention from around the world due to demands in the public security and information security sectors. Face recognition has significant advantages over other biometric modalities, and hence has tremendous application prospects. Not only so, research on face recognition presents great academic significance in image pattern recognition. This thesis is aimed to develop better face recognition technologies. It conducts research towards novel probability models and new methods based on local features using statistical image analysis and pattern recognition theories and techniques. Global decision of face classification is performed from local feature based classifiers. The thesis contains new research in the following three directions: (1) Face modeling and recognition methods based on local feature and Markov random fields; (2) Heterogeneous face synthesis and recognition based on face analogy and local features; (3) multi-modal biometric fusion of local feature based biometric recognition results. The contributions of the thesis are the following: 1. It proposes a Markov Random Field (MRF) based face recognition model, in the framework of Markov Random Field and Bayesian decision theories. There, constraints on local features as well as contextual relationships between them are explored and encoded into a cost function. The proposed MRF method, validated by experiments, provides a new perspective for modeling the face recognition problem. 2. Inspired by image texture analysis and synthesis, it proposes a new conception of “face analogy”. By local normalization and face analogy, Near Infrared (NIR) face images can be converted into VIS (Visual) face images. Thereby, heterogeneous face recognition problem can be solved using homogeneous face matching methods. This provides a novel approach for heterogenous face recognition. 3. It introduces an effective multi-modal recognition algorithm by fusing NIR face, VIS face and iris images, in which nonlinear score level fusion rules are developed. Improved performance is obtained by experiments. Finally, a NIR-VIS multi-modal face recognition system, composed of hardware and software, is developed for practical applications.
修改评论