The locations of the mouth and the two eye regions give a fairly stable measure of the size of a face. We can use these feature points to compute EM, the distance from the mid-point between the eyes to the mouth. From sample measurements, the value of EM was found to be reliable enough to predict the radius of the iris of a subject's face. The iris radius is typically expected to fall between 5% and 15% of the value of EM.
Consequently, the real-time symmetry transform is reset to use annular sampling regions that cover a radius between 5% and 15% of the value of EM. The transform is then utilized to compute two small interest maps around the previously located eye regions. The peaks of these interest maps indicate possible positions of the left and right iris. Figure shows the resulting iris loci.
The search space for the iris is centered around the old eye locus and is a square with length 25% on each side. Figure shows the actual search space windows with the best iris position represented with a + symbol. The strongest interest peak in that window is used as the iris position. The interest map is thresholded to discard all peaks which trigger symmetry values below 25% of the maximum possible output. This extra threshold allows us to avoid triggering the iris finder with other structures such as the eye brows. These and other objects generate a weak response and the iris finder might erroneously converge to their loci if the iris is not clearly visible (i.e. subject is squinting). Thus, the threshold allows us to report the absence of an iris in the search space if the peak response is too weak. Consequently, no valid peaks in the interest maps are found, and the iris localization function can merely default to the previously calculated position of the eye region. Therefore, if the individual in the image is squinting or the eyes are not clearly visible, we use the large, coarse eye-blob detection output instead of the iris finder as the position of the iris.