Orientation template matching for face localization in complex visual scenes

An important topic in face recognition as well as in video coding or multi-modal human machine interfaces is the automatic localization of faces or head-and-shoulder regions in visual scenes. The algorithms therefore should be computationally efficient and robust against distortions like varying lighting conditions. This paper describes a novel method for segmenting frontal head and shoulder views of persons from grey level images. The segmentation is done by oriented template correlation. This matching method only depends on edge information, especially the orientation of the edges. We describe the generation of a statistical model of edge orientation within a human face from a training sample. In the matching stage we calculate the probability for a face at the current image position using this model. The detection capabilities of the presented algorithm are evaluated on a large database of 1114 images each containing one or more faces.