Cascaded Face Detector with Multiple Templates

This paper proposes a novel face detection algorithm which extracts a local image structure (LIS) feature and adopts a boosting approach to construct a cascaded face detector. Due to the locality of LIS, the extracted feature is not only robust to lighting variation but also is invariant to small degrees of rotation. With this robust property a multiple-template cascaded detection algorithm has been developed which can avoid rotating the image and also keep the ability to detect slanted faces. Because the multiple templates are constructed in the initialization stage, the proposed face detector can be executed very fast. Experiments on the BioID face database have shown the efficiency of this method.