Optical clustering method for locating facial features

The localization on facial features is needed for face recognition since it helps keeping accordance between face images and building face model. In this paper, a novel method for locating facial features was presented which included two steps: filtering and clustering. Face images were firstly processed by Gabor filter into magnitude responses. In the responses, facial features demonstrated relatively high magnitude responses than other facial parts, such as cheek and forehead. By reserving high magnitude responses and removing low magnitude responses, the pixel points belonging to facial features were collected. The method adopted a clustering approach-k-means for separating pixel points into different clusters. Each cluster represented a facial feature. By testing on the ORL face database, the method shows its accuracy and speed on locating facial features, such as eyes, nose and mouth. It also exhibits high robustness in locating features on faces which have thick beard or mustache.