Eye feature extraction using K-means clustering for low illumination and iris color variety

This paper presents an approach for locating eye features in color images based on the unsupervised K-means clustering. Given the assumption that the input is an eye window containing a single eye, the proposed method detects the iris by unsupervised K-means clustering on the feature spaces of compensated red and green color channels. The iris circle is then refined using the gradient information and circular Hough transform. For the sclera detection, the r-g and r-b are utilized as they show the discriminant feature of sciera regardless of light condition and iris color. The sciera is then extended to fit the eyelids by a region growing scheme. Experiments on a collection of eye images extracted from FERET facial database and our self-collected images show a promising performance toward the low illumination and iris color variety.