The experimental database, described below, consists of training, test and difficult test images. To judge the accuracy of normalisation eye location is assessed. Since the eye can swivel and close, the centre of the eye socket is used as a fixed reference point.
|Eye location accuracy|
|Set of faces||% within 5 pixels|
Figure 1: The top left image is a normalised difficult test face. It is correctly matched to the training face on its right which has also been normalised. Below each face is the corresponding orientation image. These display the orientation of the gradient vector. The images are darkest for and brightest for . If the gradient magnitude is zero, the orientation can not be found and, for this display, the pixel is set to the lowest intensity. The grid is for n=11
Figure 2: The face in the top left is a normal test image. It is correctly matched to the training face on the right. Both faces have been normalised. Note the stability of the orientation images.