Locating facial features using threshold images

This paper describes a scheme for locating the eyes and the mouth on a facial image with slightly constrained background. The presented method reasonably assumes that only the facial image should be thresholded well enough that the hair and eyes can be distinctly separated from the cheeks. It first searches for all potential circles with specific radius as the candidates of the irises by performing the Hough transform on the binarized image, then sifts the candidates to detect the two eyes. It then roughly binds the mouth region according to the anthropometric standards and finally utilizes the integral projections to obtain the position of the mouth. The method has been tested on 223 face pictures with different illuminating conditions and varied head poses, and showed a quite satisfactory performance in both speed and accuracy