Multi-view face detection with the multi-resolution MPP classifiers

Detecting multi-view faces is a challenging task, not only because of the face variations in scale, illumination, and expression, but also of the variations caused by multiple views. In this paper, we proposed a multi-layer cascaded architecture, which can focus attention on the promising regions of the image. The whole classifiers are only evaluated on the face like parts of the image, while the most amount of background blocks are excluded by the first few layers of the detector. Instead of using predefined priori knowledge about face view partition, we divide the sample space automatically by the branching competitive learning network at different discriminative resolutions. To maintain the high detection efficiency, we adopt the simplified Support Vector Machines (SVMs), called the mirror pair of points (MPP) classifiers, as the component of our detection system. Experimental results show that our system is competitive with other systems presented recently in the literature.