An Intelligent and Learning-Based Method for Automatic Iris Segmentation

Iris segmentation is one of the most important pre-processing stage for an iris recognition system. Thequality of iris segmentation results dictates the iris recognition performance. The main difficulty about irissegmentation lies in the heterogeneity of the captured iris images. For example, different iris acquisitiondevices use optical sensors that have different optoelectronic responses to photons, resulting in substantialdeviation in pixel intensity in regions of the iris, the pupil and the sclera. Besides, the occurrence of irisocclusion by eyelids or eyelashes is a stochastic process. When we compare eye images across databasescollected by different academic units at different time, such indefiniteness become even more serious. All ofthese factors which seriously affect the segmentation results are very hard to be predicted by any feature beforesegmentation starts.In the past, methods of either learning-based (for example, neural network) or non-learning-based (forexample, Hough Transform) have been proposed to deal with this topic. However, there does not exist anobjective and quantitative figure of merit in terms of quality assessment for iris segmentation (to judge whethera segmentation hypothesis is accurate or not). Most existing works evaluated their iris segmentation quality byhuman. In this work, we propose KIRD, a mechanism to fairly judge the correctness of iris segmentationhypotheses. KIRD is a method that gives a numerical assessment to the quality of the iris segmentation result.The key concept of KIRD is inspired by observing the substantial difference of the pixel values inside andoutside the segmentation contour. By pre-processing the iris images with PCA and K-means, then integratingpixel value difference along a pre-defined segment on a given iris segmentation hypothesis, KIRD value canbe computed. In our preliminary experiment, it shows KIRD is a reliable metric to give an objective assessmentfor the quality of iris segmentation hypothesis.On the foundation of KIRD, we propose AILIS, which is an adaptive and iterative learning method foriris segmentation. Since all of the important factors that affect the iris segmentation results are stochastic, theproposed algorithm tries to learn the underlying distribution using multi-model approach, with a “divide-andconquer”fashion. In AILIS, a mechanism is built in order to distinguish the wrong segmentation results fromthe correct ones. Given such ability, AILIS is able to learn a new model from the past to describe the appearanceof iris images which are unseen before and able to learn from past experience and automatically build machinelearningmodels for iris segmentation for both gray-scale and colored iris images.