Abstract
Face detection is a more and more important topic nowadays. Because of the need on machine vision and surveillance system, we have to let a computer detect the faces in a picture rapidly and accurately. Viola and Jones [1] proposed a rapid and robust method, but it is just suitable for a frontal and upright face. Detecting multi-view faces is also a discussed problem today. In this thesis, a face detection system applicable for wider range is proposed. It contains analyzing the face pose, rotating the face back, and applying the traditional viola-jones detector. First, we search for potential face regions by performing skin-color detection. From these skin-color blocks, check if there is a pair of eyes inside and preserve window candidates that are much like faces. Second, according to the result of the pair of eyes candidate, we choose the angle to rotate each face candidate to frontal and upright face. Finally, the rotated face candidate is computed by viola-jones detector to judge whether it is a genuine face or not. In the experiment, our method extends the detection range of rotation-off-plane (ROP) and rotation-in-plane (RIP) face. Some faces, including profiled and rotated faces, that can not be detected by original viola-jones face detector are detected by our modified face detector with complemented pre-processing. The processing time is also acceptable. Especially for simple and small size images, it can achieve nearly real-time.