Iris recognition based on non-local comparisons

Iris recognition provides a reliable method for personal identification. Inspired by recent achievements in the field of visual neuroscience, we encode the non-local image comparisons qualitatively for iris recognition. In this scheme, each bit iris code corresponds to the sign of an inequality across several distant image regions. Compared with local ordinal measures, the relationships of dissociated multi-pole are more informative and robust against intraclass variations. Thus non-local ordinal measures are more suited for iris recognition. In our early work, we have built a general framework "robust encoding of local ordinal measures" to unify several top iris recognition algorithms. Therefore the results reported in this paper improve state-of-the-art iris recognition performance essentially as well as evolve the framework from pair-wise local ordinal relationship to non-local ordinal feature of multiple regions. Our ideas are proved on CASIA iris image database.