Extending Concentricity Analysis by Deformable Templates for Improved Eye Extraction
David E. Benn, Mark S. Nixon, J.N. Carter · ePrints Soton (University of Southampton) · 1999
We present a new two stage approach to extracting human eyes from face image by structure alone, without priming on likely position. First a concentricity map of the face is created; the eyes are usually amongst the highest peaks of this map. Next, the energy of a deformable eye template, initialised at concentricity peaks, is optimised using a genetic algorithm. The two highest energy locations are deemed to contain the eyes. We have improved the robustness of an earlier concentricity map and improved Yuille et al's eye template model by removing two parameters whilst also removing the template's dependency on internal energies. Our enhancements has been verified on a database of one thousand full frontal face images, yielding a success rate just exceeding 90% correct eye extraction.