Model interpolation for eye localization using the discriminative generalized Hough transform

Abstract

The Discriminative Generalized Hough Transform (DGHT) is a general method for the localization of arbitrary objects with well-defined shape, which has been successfully applied in medical image processing. In this contribution, the framework is used for eye localization in the public PUT face database. The DGHT combines the Generalized Hough Transform (GHT) with a discriminative training procedure to generate GHT shape models with individual positive and negative model point weights. Based on a set of training images with annotated target points, the individual votes of model points in the Hough space are combined in a maximum-entropy probability distribution and the free parameters are optimized with respect to the training error rate. The estimated model point specific weights reflect the important model structures to distinguish the target object from other confusable image parts. Additionally, the point weights allow for a determination of irrelevant parts in the model, which can be eliminated to make space for new model point candidates from training images with high localization error. The iterative training procedure of weight estimation, point elimination, testing on training images, and incorporation of new model point candidates is repeated until a stopping criterion is reached. Furthermore, the DGHT framework incorporates a multi-level approach, in which the searched region is reduced in 6 zooming steps, using individually trained shape models. In order to further enhance the robustness of the method, the DGHT framework is, for the first time, extended by a linear model interpolation for the trained left and right eye model. An evaluation on the PUT face database has shown a success rate of 99% for iris detection in frontal-view images and 97% if the test set contains a large head pose variability.