Accurate And Real-Time Face Detection Based On Near-Infrared Image

Face detection is a critical stage of face recognition system, with the task of indicating whether there exist faces on a given image and localizing the face region. In recent years, it also has wide applications in areas such as content-based image retrieval, crowd surveillance, and intelligent human-computer interface. Face detection is a complex pattern recognition issue, the complexity consist of variety of face pose, uncontrolled environmental illumination, and accessory changing.In order to reduce the influence of illumination, a lot of researches have been done intensively. Compared with finding illumination insensitive face representation features and robust algorithm, using active near-infrared (NIR) imaging system is a more effective way.This thesis proposes a novel accurate and fast face detection algorithm based on NIR image. The proposed face detection algorithm consists of face candidate region detection, and accurate eye localization. Initially, the Haar feature and AdaBoost algorithm is introduced here to detect face, and then detection of the presence of glasses is performed in the candidate face area. If glasses do not present in face area, N-Quoit filter based on mathematical morphology will be used to locate eyes utilizing the "bright pupil" effect in NIR image. Otherwise, eye location can be determined by the "Faceness" criterion.The main contributions of the thesis are summarized in the following. (1) First is the introduction of a new eye localization method based on N-Quoit filter utilizing the "bright pupil" effect in NIR image. (2) Second, an accurate and fast glasses detection method is presented, which is capable of separating the facial image with glasses from that without glasses. (3) Third is the introduction the "Faceness" criterion, which is used for eye localization.This thesis implements the AdaBoost trainer and face detector. Experimental results demonstrate that the proposed face detection algorithm is capable of achieving high accuracy and real-time performance, and is an effective solution to the issue of specular reflections of NIR lights on glasses in eye localization.