Hybrid neural method for locating eyes in facial images

Abstract

Locating facial features automatically in a scene is an essential but relatively unsolved problem. We develop a novel hybrid neural method for human eye location by the following stages: (1) facial images are preprocessed for normalization; (2) the candidate regions of eyes are detected using an improved version of radial basis function (RBF) networks, which is capable of dealing with those facial images containing some changes of eye size or face orientation; and (3) a hierarchical knowledge-based approach is presented to evaluate these candidates and locate the positions of both eyes. The proposed method is tested on various face images. The experimental results illustrate the effectiveness of our method.