Feature-Based Affine-Invariant Detection and Localization of Faces.

The accuracy of human face detection and localization in images and video is a crucial factor influencing the performance of biometric face authentication and recognition systems. Recently this subject attracted a lot of attention by researchers and companies and its applications emerged in various areas including surveillance, security, and computer games. This thesis describes a novel person-independent method for finding and localizing faces in authentication scenarios. Such scenarios involve situations where a person stands or sits in front of a camera in order to gain access. The objective was to develop an algorithm which uses only still grey-level images, copes well in the presence of cluttered background and accurately localizes faces including eye centres. Many of the methods that have been reported in the literature only partially fulfil these requirements, in particular, a few methods focus on precise eye localization. To address these issues, we propose a novel bottom-up face detection and localization algorithm which exploits statistical feature detectors as the means of image capture effects removal. Our method uses both, a constellation (shape) model and shape-free texture model to select the best face location hypothesis among multiple hypotheses generated by the feature detectors. The constellation model utilizes a distribution of the transformation from a proposed model space into the image space. The texture (appearance) model is based on a cascaded Support Vector Machine classification. Both, an extensive analysis and a performance evaluation on several realistic face databases will be discussed in this thesis. We show that by utilizing the proposed verification of hypotheses, a significant performance boost is achieved compared to the performance of feature detectors alone