Using artificial neural networks and feature saliency to identify iris measurements that contain the most discriminatory information for iris segmentation

One of the basic challenges to robust iris recognition is iris segmentation. To represent the iris, some researchers fit circles, ellipses or active contours to the boundary pixels of the segmented iris. In order to get an accurate fit, the iris boundary must first be accurately identified. Some segmentation methods operate on a pre-processed gray-scaled image, while others use a thresholded binary edge image. The Hough transform is a popular method used to search for circular or elliptical patterns within the image. Many irises are slightly elliptical, and suffer from eyelid/eyelash occlusion, specular reflections and often the pupil and iris centers are not co-located. Each of these issues can cause a segmentation error. This research uses of a feature saliency algorithm to identify which measurements, used in common iris segmentation methods, jointly contain the most discriminatory information for identify the iris boundary. Once this feature set is identified, an artificial neural network is used to near-optimally combine the segmentation measurements to better localize and identify boundary pixels of the iris. In this approach, no assumption of circularity is assumed when identifying the iris boundary. 322 measurements were tested and eight were found to contain discriminatory information that can assist in identifying the iris boundary. For occluded images, the iris masks created by the neural network were consistently more accurate than the truth mask created using the circular iris boundary assumption.