Combination of facial landmarks for robust eye localization using the Discriminative Generalized Hough Transform

The Discriminative Generalized Hough Transform (DGHT) is a general and robust automated object localization method, which has been shown to achieve state-of-the-art success rates in different application areas like medical image analysis and person localization. In this contribution the framework is enhanced by a novel fa-ciallandmark combination technique which is theoretically introduced and evaluated for an eye localization task on a public database. The technique applies individually trained DGHT models for the localization of different facial landmarks, combines the obtained Hough spaces into a 3D feature matrix and applies a specifically trained higher-level DGHT model for the final localization based on the given features. In addition to that, the framework is further improved by a task-specific multi-level approach which adjusts the zooming-in strategy with respect to relevant structures and confusable objects. With the new system it was possible to increase the iris localization rate from 96.6% to 97.9% on 3830 evaluation images. This result is promising, since the variation of the head pose in the database is quite large and the applied error measure considers the worst of a left and right eye localization attempt.