Precisely locating human eyes from front-view face images

Abstract

In this paper, a novel method for automatic localization of human eyes from front-view face images is presented. By applying the grayscale morphology operation to a face image, we can obtain a valley map with a good property in which as a whole the number of pixels at a grayscale decreases significantly as the grayscale level increases. The valley map is then binarized with a set of threshold values determined adaptively, and several candidate eye locations are detected from the resultant binary images. The final eye positions are determined using a Principle Component Analysis (PCA) method. Experimental results on the AR, Yale and our own face image databases show that the correct eye location rate is over 93% and the location disparity on average is below 1/3 of the radius of an eye ball.