4 Conclusions

  The recognition experiments performed on the M2VTS database show (see Table  1 ) that the optimised robust correlation outperformed methods based on the dynamic link architecture [ 6 , 2 ]. The speed of the method is adequate for the identification scenario. Moreover, the near real-time response allows repeated application of the method to increase the reliability of the reject/accept decision, especially in border-line cases. The performance of the method depends neither on pre-registration of the images to be matched, nor on the success of feature detectors for localisation. On the contrary, registration of the compared images is achieved as an integral part of the identification process. The method is robust and needs no manually built models.

The method seems promising as implemented, but, in our opinion, it can be further strengthened by the following two improvements. At present, the same sampling density is used throughout the image. Non-uniform sampling controlled by the discriminative power of the different regions of the face is likely to improve the recognition performance [ 7 ]. Secondly, inclusion of colour information in the proposed method is fairly straightforward. Given a metric for measuring the distance between two pixels in colour space the optimised robust correlation can be applied directly without any modifications.