Enhancement of template-based face detection by belief propagation in ordered component search

Abstract

Face detection problem is a high-dimensional problem. Here, we proposed an enhancement to face detection approaches, particularly, template-based approaches, in which we avoid the exhaustive search while we obtain face scale, rotation angle, and facial component locations. Our biologically inspired algorithm quickly deduces face position, scale, and orientation, as well as location of facial feature, by an ordered search and belief propagation. Our approach has several advantages over image-based and feature-based face detection methods. In addition, our approach can be used to detect both real faces and non-real faces including symbolic faces, cartoons, puppets, and caricatures. We used our approach to enhance Viola-Jones face detector in terms of running time and accuracy. We presented results of our tests on Multi-PIE dataset which show decrement of the running time from 145.11 seconds to 1.48 seconds and increment the detection accuracy from 67.4% to 93.8%. © 2012 Springer-Verlag GmbH.