Facial Feature Detection Using Generalized LVQ and Facial Shape Model

A method for detecting the facial feature points, such as the pupil, subnasal point, and corners of the mouth, is proposed. The proposed method is composed of two stages: candidate detection of facial feature points and optimization of these points by using a facial shape model. The candidates for each facial-feature-point are extracted from a face image by using generalized learning vector quantization classifiers, and the most suitable facial feature points are then selected from the facial feature point candidates obtained in the first stage. The facial shape model is utilized to constrain the alignment of facial feature points while abnormal candidates are estimated by the least-median-of-squares method. Experiments using a large still-face dataset with various illumination conditions demonstrate that the proposed method can extract facial features precisely under varying illumination and facial-expression conditions.