Detection of eyes from human faces by Hough transform and separability filter

Abstract

Detection of the eyes in human facial images is important in preprocessing for face recognition, since the results can be used to determine the position, size, and direction of the face. This paper describes a method of detection of both irises from a gray-scale facial image. In this method, first a small region ("blob") which is likely to be an iris is determined by using a separability filter (developed by Fukui and Yamaguchi). Second, taking all the blobs, the costs of all segments connecting the irises and the inclinations of the segments which are within a predetermined value are computed. This cost can be calculated from the regions surrounding the blobs, by using the number of votes for the maximum-voting circle, the separability of the surrounding small regions, the mean brightness within the iris, and the normalized correlation coefficient between a partial image including a pair of blobs and the template for the eyes. The pair of blobs that minimizes the cost is regarded as the irises. The proposed method has been tested by using all 150 images (excluding images with glasses) from the University of Bern Facial Image Database. The experimental results show that the success rates are between 95.3% (worst) and 97.3% (best). If minor errors are tolerated, the success rates are between 96.7% (worst) and 99.3% (best). © 2005 Wiley Periodicals, Inc.