A method of feature detection for human face images is proposed. One of the remarkable features on human faces is "eye", which has a symetrical structure in image intensities. Some symmetrical transforms on image intensities are presented for detecting face features such as eye, nose or mouth. Image areas centered around the detected features are mapped to log-polar coordinates and then processed by auto-correlation to obtain rotation and scale invariance. Finally the auto-correlation is classified in the feature space by primary component analysis. Experimental results are also presented.