Automatic extraction of human facial features

This paper addresses a method of extracting human facial features from the head-and-shoulder images used in videophone communications. A basic idea behind the scheme is to make full use of the information in both the temporal and spatial domains to detect facial features. The new approach to facial feature extraction consists of a temporal and a spatial eye-mouth finder (EMF). The temporal technique eliminates the influence of background and gives a rough and reliable position estimation of the interesting objects, e.g. eyes and mouth, and the latter reaches an accurate result which is needed in videophone communications. Another characteristic of this method is its simplicity. Some basic operators can be used twice, i.e., they are used not only in the temporal domain but also in the spatial domain. The results on the test image sequences are satisfactory.