Conclusion

We have shown how a well-known method of two-dimensional face recognition can be applied to three-dimensional face models achieving reasonably low error rates, depending on the surface representation used. Drawing on previous work combing face recognition eigenspaces [11], we have applied the same principle to multiple threedimensional face recognition systems, showing that the combination method is
applicable to both two-dimensional and three-dimensional data. Using FLD as an analysis tool, we have confirmed the hypothesis that although some surface representations may not perform well when used for recognition, they may harbor highly discriminatory components that could complement other surface spaces. Iteratively improving error rates on a small test set, we have built up a combination of dimensions extracted from a variety of surface spaces, each utilising a different surface representation. This method of combination has been shown to be most effective when used with the cosine distance metric, in which a selection of 184 dimensions were combined from 16 of the 17 surface spaces, reducing the EER from 11.6% to 8.2%. Applying the same combined surface space to an unseen test set of data presenting typical difficulties when performing recognition, we have demonstrated a similar reduction in error from 11.5% to 9.3% EER.

Evaluating the combined system at its fundamental level, using 1,079,715 verification operations between three-dimensional facial surfaces, demonstrates that combining multiple surface space dimensions improves effectiveness of the core recognition algorithm. Error rates have been significantly reduced to state-of-the-art levels, when evaluated on a difficult test set including variations in expression and orientation. However, we have not applied any additional heuristics, typically incorporated into fully functional commercial and industrial systems. For example, we have not experimented with multiple facial alignments, optimising crop regions or storing multiple gallery images. All of which are known to improve error rates and can easily be applied to the combined systems presented in this paper. With these additional measures in place, it is likely that the improvements made to the core algorithm will
propagate through to producing a highly effective face recognition system. Given the fast 3D capture method, small face-keys of 184 vector elements (allowing extremely fast comparisons), invariance to lighting conditions and facial orientation, this system is particularly suited to security and surveillance applications.