Robust Face Recognition based on Color and Depth Information

One of the most important advantages of automatic human face recognition is its nonintrusiveness property. Face images can sometime be acquired without user's knowledge or explicit cooperation. However, face images acquired in an uncontrolled environment can appear with varying imaging conditions. Traditionally, researchers focus on tackling this problem using 2D gray-scale images due to the wide availability of 2D cameras and the low processing and storage cost of gray-scale data. Nevertheless, face recognition can not be performed reliably with 2D gray-scale data due to insu_cient information and its high sensitivity to pose, expression and illumination variations. Recent rapid development in hardware makes acquisition and processing of color and 3D data feasible. This thesis aims to improve face recognition accuracy and robustness using color and 3D information.In terms of color information usage, this thesis proposes several improvements over existing approaches. Firstly, the Block-wise Discriminant Color Space is proposed, which learns the discriminative color space based on local patches of a human face image instead of the holistic image, as human faces display different colors in different parts. Secondly, observing that most of the existing color spaces consist of at most three color components, while complementary information can be found in multiple color components across multiple color spaces and therefore the Multiple Color Fusion model is proposed to search and utilize multiple color components effectively. Lastly, two robust color face recognition algorithms are proposed. The Color Sparse Coding method can deal with face images with noise and occlusion. The Multi-linear Color Tensor Discriminant method harnesses multi-linear technique to handle non-linear data. Experiments show that all the proposed methods outperform their existing competitors.In terms of 3D information utilization, this thesis investigates the feasibility of face recognition using Kinect. Unlike traditional 3D scanners which are too slow in speed and too expensive in cost for broad face recognition applications, Kinect trades data quality for high speed and low cost. An algorithm is proposed to show that Kinect data can be used for face recognition despite its noisy nature. In order to fully utilize Kinect data, a more sophisticated RGB-D face recognition algorithm is developed which harnesses theColor Sparse Coding framework and 3D information to perform accurate face recognition robustly even under simultaneous varying conditions of poses, illuminations, expressionsand disguises