6 Discussion

The difficult test images are less accurately normalised than the training or normal test images. This is expected due to the greater noise in the difficult test images.

The descriptors used for the templates are selected for their low variation with respect to illumination. If grey-levels are ranked locally the resulting pattern is less affected by global variation; for example the cheeks are generally local maxima with regard to intensity, even if one side of the face is in shadow. The direction of the gradient vector, irrespective of magnitude, also captures facial anatomy despite changing lighting conditions. - figure 2.

The direction of the gradient vector contributes to 50% of the normalisation method and is the sole descriptor for recognition. In contrast many studies have employed grey-level values as descriptors. The orientation histograms yield high recognition rates. Figure 2 illustrates how orientation images for the same person can be stable.

The recognition results for the normal test and difficult test images, using all windows, vary little between n =7 and n =15. Closer inspection of the results reveals two trends. Some faces recognised at lower values of n are not recognised as n increases. This is because they are not normalised with a sufficient degree of accuracy for the spatial detail demanded. However, other faces, which are more accurately normalised, are only discriminated at higher values of n . This turnover of faces indicates that the normalisation method is constraining the recognition method. We have performed further experiments on the normal test faces and improved the normalisation. The recognition rate, using all windows, rises to the level of the best results above for subsets of windows.

The combined templates are more effective than any of the component templates. However each of the component templates performs effectively alone.

The template based on gradient magnitudes does not prove as successful as that based on ranked gradient magnitudes. Ranking may be more effective because it does not rely on absolute or relative values, but only on the local order of values. This may make it less specific, but more able to cope with variation [ 3 ].

The use of the subset of windows for obscured images appears to enhance the results. This is a simple way of removing expected noise; if a test image window is obscured, it is more likely to produce a worse match with the corresponding training image window. If only the better matching windows are added, such noise might be omitted. However the results for the normal test images were also improved. Moreover recognition rates approaching 100% were obtained even though only 85% of normal test face eyes were located within five pixels. There are a number of possible reasons for this:

1. All normal test faces were normalised to a reasonable extent; none were missed.
2. For some people in the database, there may be a systematic eye location error. This would depend on the fit between a particular face and the template, produced by averaging. Thus although eye location could be improved, good recognition is still possible.
3. The normal test images often differ from the training images in ways which would promote the use of a limited number of windows. Such changes include different hairstyles, donning glasses, altering expression, rotation and extensive lighting glare.
4. If one eye is more accurately positioned than the other , windows on one side of the face might return better matches.

It is also noteworthy that a high recognition performance is maintained with just 12.5% of the windows. This indicates that the description is accurate on a local level and doesn't rely on a global comparison.