Multi-pose Face Detection And Recognition On Nonlinear Manifold

Face detection and recognition is the theory on locating the face or facial feature and recognizing the identity, expression and pose et al. in the image or video. Face detection and recognition is hard due to the variation of the face pattern caused by different types of variations in face images, such as pose, illumination and expression. The existing algorithms have good performance under controlled environment, but without enough adaptivity for the variations mentioned above.Among all kinds of variations, pose variation is the hardest one to deal with and therefore contributes most to the error of the detection and recognition algorithms. Therefore, this thesis aims to study the fundamental theory and key technology of multi-pose face detection and recognition such as multi-pose face detection, facial feature localization, pose estimation and multi-pose face recognition. At the same time, the nonlinear manifold learning algorithm is applied or improved to solve the problems such as pose estimation and face recognition. The major research works and contributions of the thesis include:(1) As for frontal face detection, a Real Adaboost algorithm and region edge-orientation field matching method is proposed. The edge-orientation field is extracted form the original face images, and Real Adaboost algorithm is applied to obtain the face pattern in an iterative way. The cascaded classifier structure and multi-resolution image research strategy are also adopted to guarantee the real time performance.(2) A new algorithm called Co-Adaboost is developed for multi-feature multi-pose face detection. The edge-orientation field feature and Harr-like feature are both applied, and then merged by using Co-Adaboost algorithm. The whole pose angle is decomposed coarse-to-fine, and a classifier-pyramid architecture is adopted to detect multi-pose faces efficiently.(3) In order to solve the problem of facial feature localization under large-scale pose angle, a hierarchical edge orientation field matching based algorithm is proposed. A new measurement coined structural Hausdoff distance is developed for edge orientation field matching. The global edge orientation field matching is applied to estimate the coarse position of the features, and then the further local feature edge orientation field matching is carried out for precise localization.(4) An Isomap-based nonlinear manifold learning algorithm is proposed for pose estimation. The structural Hausdoff distance is employed for edge orientation field matching, and feature adjusting is carried out according to the feature positions. Isomap is applied to map the input data points into the dimensionality-reduced space, thereafter a pose parameter map is learned, and the pose of the new facial image could be estimated in this way.(5) A correlative sub-region mapping method is proposed to recognize the multi-pose face images. The face is divided into several sub-regions, and the influence to image caused by the pose variation is decomposed into shape mapping and texture mapping between correlative sub-regions. A new technique coined two-dimensional coupled component analysis is developed to construct these mapping functions. The discriminative power of each sub-region is estimated using Bayesian framework, and thereafter the final recognition result is obtained by combining the measures in all sub-regions.(6) A new discriminant analysis algorithm on nonlinear manifold is proposed and applied to face recognition. Using geodesic distance to discover the intrinsic geometry of the manifold, the geodesic Gabriel graph is constructed to locate the critical local regions where the local linear discriminants would be learned. The global nonlinear discriminant is achieved by merging the multiple local discriminants. The soft margin criterion based merging algorithm assigns the best weight to each local discriminant in an iterative way and upgrads the detection accuracy stepwise.