Whole face templates based on the above descriptors are effective for normalisation. If the templates are combined, performance is improved. If normalisation could be made more accurate, recognition would benefit.
The normalisation method is robust enough for complete and obscured faces with varying backgrounds. The addition of further descriptors to the analysis would be expected to enhance the performance.
The recognition method presented is based on the orientation of the gradient vector. Further investigations are required to explore the characteristics of this descriptor. It appears to be stable with changes in illumination and yields distinctive descriptions for recognition.
Recognition performance can be enhanced by selecting a subset of closely matching windows. Whilst it is clear how this can omit obvious noise, the method also has a significant effect on the normal test images. The main reasons for this boost need to be determined.
The recognition scheme could be developed by employing several descriptors.