Multi-Pose Face Recognition

Face recognition is one of the key techniques in computer vision, image processing, pattern recognition and artificial intelligence, and multi-pose face recognition is a nontrivial task to tackle in this domain. This thesis focuses on multi-pose face recognition, facial components are localized based on ASM (Active Shape Model) and the face sequence matching is implemented based on MSM (Mutual Subspace Method).ASM is a deformable model and the control parameters can be obtained by data learning. Tuning these parameters within some ranges, the target shapes can be altered. Normally the configuration and contours of facial components in multi-pose face images changes greatly, this thesis presents a method to localize them with ASM. Through data alignment, statistic analysis and PCA (Principal Component Analysis) on the training data set, the parameters of ASM are obtained. Using the parameters self-adapting algorithm and Local Structure Model, the positions of facial components can be estimated in the face images of different poses. The face candidate regions are normalized for multi-pose face recognition. An experiment is conducted on 150 face images, and the results are satisfied. The average localization error of eye center is about 4 pixels and the average localization errors of nose tips and mouth center are a little bigger, basically within 10 pixels. It shows with this method facial components can be well localized.This thesis studies multi-pose face recognition method based on MSM as well. After face normalization, feature extraction and sequences forming, the subspace of sequences is computed. The face is recognized by computing the cosine value of the angle between the input and the reference subspace. An experiment is conducted on 100 face sequences, the recognition accuracy is 93.8%, which demonstrates that the multi-pose face recognition method based on MSM can improve the recognition accuracy of multi-pose faces.