Periocular and iris local descriptors for identity verification in mobile applications

Abstract

The 2016 International Conference on Pattern Recognition hosted the MICHE-II Contest, with the aim at biometric identification on mobile devices. This paper describes the ideas behind one of the contest submissions, in particular the one ranked 6th (4th in cross-device), including different novelties in relation to the original proposal. Our approach is based on the extraction of local descriptors previous to classification. In this sense, starting from a common iris segmentation information, different normalization procedures are considered, analyzing the use of both iris and periocular patterns. A collection of local descriptors is computed on those patterns, evaluating their performance by means of different classification paradigms in a 10-fold cross validation experiment. Our results suggest the great utility of the periocular area for this problem and dataset. Finally, periocular based classifiers are evaluated on the test set, evidencing an improvement in relation to our original submission, with a promising close future improvement if a fusion approach is adopted.