Robust Feature Extraction for Facial Image Quality Assessment

초록

With the aim of developing a fully automatic system to verify if facial images meet ISO/IEC 19794-5 standards and ICAO requirements, this paper proposes a robust method to detect facial features including two eye centers and four lip features. The proposed method takes skin color information into account to restrict the areas where facial features are observed. In order to overcome the in uence of head poses and illumination, which are the main factors of unsuccessful eye detection, face images are adaptively thresholded by considering the distribution of facial skin. And then, using circular lters, which are specially designed to nd circular shapes in images, two eye centers are detected independently. To accurately detect lip feature positions regardless the presence of beard or mustache, the proposed method fuses edge and color information together. An experiment was performed on a subset of the FERET database, and the experimental results demonstrated the accuracy and robustness of the proposed method.