Eye Location Based on Eye Map and SIFT Features

Eye alignment is the precondition for face registration, and the accurate eye localization is the main method for eye alignment. In this paper, a novel eye localization method based on eye map and SIFT feature is presented. First, based on the fact that eyeball is dark and round; those pixels darker than their surroundings are those from a face image. Then, connected regions in the eye map and their geometric centers are obtained by a rank order filter. Once more, candidate suitable eyeball pairs are selected based on a set of geometric constraints. And finally, SIFT features are extracted from the area of candidate eyeball pairs, with which the corresponding values are obtained by the support vector machine regressor. The pixel corresponding to the maximal value is just the accurate eye localization expected. Experiments show that the method not only has higher location accuracy but also has faster computation speed.