Human face recognition by computer.

Two computer program systems are designed to recognize a human face automatically, one system is concerned with recognition of persons from their side profile image while the other is concerned with recognition from frontal view. Each system is consist from two module: module 1 for facial feature extraction and module 2 for identification of faces . For the first system, photographs of faces' profiles when the persons are locking straight ahead are produced for 6 persons by using back lighting. Six interesting points (Nose points, nose bottom point, mouth point, chin point, forehead point, and eye point) are extracted and used to locate the profile outline automatically. Five curve segments were obtained from these six points, then from the curve segments the system extracted 26 features for every person. This process is repeated twice for the same person's head tilted slightly upward and slightly downward. The mean and the stander deviation for each feature are calculated from the 3 posess and stored in a data feature file as a reference for comparison with an unknown profile face to be f introduced to the system for the seek of recognition.
In the above processes the numerical cubic B-splines method is used to detect corners. It if found that a 100% recognition ability is achieved for straight forward posses as well as upward or downward tilt of the head by less than 5 degree. For the second system given a frontal view photograph of a face, an intensity projection of face is obtained from the thin rectangular strip starting from top of the forehead to the bottom of the chin. A program is written to compute the Walsh power spectrum coefficients of the intensity variation in the region. These coefficients which gives the profile features of a face are stored in a data feature files as a reference for comparison with any unknown face to be introduced to the system for seek of identification. The face profile region (i.e. the strip) must be located very accurately in order to get a representative profile. This task is accomplished in this,, work by locating the eyes position very carefully using Hough transform and taking the middle of the distance between the two eyes as the central line of the strip. The developed algorithms and the programs written took into consideration of the slightly changed poses of a person. It is found that the identification of the profile features obtained by the Walsh power spectrum are more sensitive to side-to-side rotation than to up and down movement. The identification module of the system may completely fail if both rotation and tilt movement occur in the same target image.