Research On Face Alignment Using ASM

Automatic Face Recognition (AFR) aims at endowing computers with the ability to identify different human beings according to face images. Such a research has both significant theoretic values and wide potential applications. After more than 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, especially the accurate facial feature location problem, which is the prerequisite for sequent feature exaction and classification. In this thesis, facial point location using statistical learning methods is studied after a recent overview of AFR research and development.Study facial feature alignment problem and provid a thorough survey of the algorithms, and then focus on face alignment using Active Shape Model. Point Distribution Model is described, and three aspects of ASM are discussed: aligning the training set, modeling shape variation and choice of number of modes. Active Appearance Model and Image Warping technique, especially the Piece-Wise affines algrithom, are described.Process of training examples collection is introduced, and key points are marked in images and models are trained using ASM for face alignment. Three models are builded for the search, including model of eyes, model of mouth, and model for the whole face. When searching a new face, we use the whole face model first, and then use the other two to adjust the result. Our experiments have illustrated the better performance on face alignment. Representation of model is discussed with the experiment using different numbers of key points and the experiment using amodel without the face outline. Average search time and point-to-point error are computed and compared. Reasons for the fail results are investigated, like changes of illumination, pose, expression and so on.Research on alignment evaluation problem, and propose a statistical learning approach for constructing an evaluation function, as the lack of convergence guide line and evaluation of ASM alignment results. The design of classifier is discussed, and the performace of classifier is decided by the selection of feature space and learning algorithm. Gabor wavelets are observed and AdaBoost learning method is introduced. Then a nonlinear classification function using Gabor feature and AdaBoost learning method is learned from a set of positive (good alignment) and negative (bad alignment) training examples to effectively distinguish between qualified and un-qualified alignment results. Experimental results demonstrate that the classification function learned using the proposed approach is effective and provides semantically more meaningful scoring.