Application of a post-processing model based on HMM for face recognition in video

In this thesis, the rarely concerned problem of data source in face recognition is investigated, and a novel post processing HMM-based solution is proposed. Data source problem is first empirically investigated through evaluating systematically the Eigenfaces sensitivity to variations of pose and illumination by Lambertian reflection model and 3D face model, which reveals that the changes of pose and illumination abruptly degrade the Eigenfaces system. This problem is explicitly defined as curse of data source for highlighting its significance. Aiming at solving this problem, combining the recognition rate with the analysis of the data sources, two methods is proposed to evaluate the overall performance of specific face recognition approach with its robustness against the low-quality data sources considered. Finally, a post-processing method is proposed to improve the robustness of the recognizer under unconstrained environment. Experimental results have impressively indicated the effectiveness of the proposed post-processing solution to tackle the curse of data source problem.