PDF《中国科学院研究生院 博士学位论文》

30 years'

development, AFR has made great progress especially in the past ten years. The state-of-the-art AFR system can perform identification successfully under well-controlled environment. However, evaluation results and practical experience have shown that AFR technologies are currently far from mature. A great number of challenges are to be solved before one can implement a robust practical AFR application. The following key issues are especially pivotal: (1) the accurate facial feature location problem, which is the prerequisite for sequent feature exaction and classification;

(2) efficient face representation and corresponding classifier with high accuracy;

(3) how to improve the robustness of AFR to inevitable mis-alignment of the facial feature. In addition, elaborate system design is also as important for developing robust and practical AFR systems. In this thesis, the above-mentioned key issues are studied, aiming at robust and practical AFR systems. And the main contribution of this thesis includes: 1. Provided a thorough survey of the AFR history and the state-of-the-art The latest AFR survey was published in the year 2000, which in fact surveyed the AFR researches before 1999. This thesis has provided a more recent overview of the AFR research and development. The survey starts from the discussion of the general computational model of AFR methods. Then, the AFR researches are partitioned to three historical stages according to the characteristics of technical mainstreams. For each stage, its technical characteristics are summarized and typical milestone methods are mentioned. After that, AFR methods are further categorized according to facial feature extraction, face representation, and classification separately. We also survey the main public face databases and performance evaluations protocols, based on which the state-of-the-arts of AFR are summarized. Main commercial systems are introduced as well with their technical sources. Finally, the challenges and technical trends in AFR fields are discussed. 2. Studied facial feature alignment problem, and proposed a Local Texture Constraint Active Appearance Model. Accurate facial feature alignment is the prerequisite of an AFR system. Active Shape Model (ASM) and Active Appearance Model (AAM) are the main methods for this problem. Abstract IV On the basis of analyzing their merits and demerits, LTC-AAM is proposed by combining the ASM and AAM. The main points of LTC-AAM include: (1) Inherit the local texture matching model in ASM and enhanced the local texture model for facial contour landmarks by exploiting explicitly the edge constraints;

(2) As in AAM, the energy function measuring the matching degree between the model texture and present texture are acted on active shape searching, therefore the shape parameter adjusting in ASM is unified into the same framework of variable optimization in AAM;

(3) Linear prediction of the appearance parameter variance using the present texture as in AAM is retained for the heuristic search in the variable optimization procedure. These points have made LTC-AAM integrate the strengths of ASM'

s local searching model, AAM'

s global texture constraints and AAM'

s heuristic parameter optimization. Therefore, our experiments have illustrated the better performance of LTC-AAM on face alignment. 3. Investigated the Gabor wavelet face representation method, and proposed a AdaGaDA-based AFR approach. Efficient face representation is the key for a good AFR method. For its eminent characteristics in spatial local feature exaction and orientation selection, Gabor-based face representation has attracted much attention. Elastic Graph Matching (EGM) and Gabor feature Fisher Clas........