Measuring the Quality of IRIS Segmentation for Improved IRIS Recognition Performance

摘要

In this paper, we present three versions of an open source software for biometric iris recognition called OSIRIS_V2, OSIRIS_V3, OSIRIS_V4 which correspond to different implementations of J. Daugman's approach. The experimental results on the database ICE2005 show that OSIRIS_V4 is the most reliable on difficult images while OSIRIS_V2 is the fastest. So, we propose a novel strategy for iris recognition using OSIRIS_V2 for good quality images and OSIRIS_V4 when the quality of the segmentation of OSIRIS_V2 is not sufficient to ensure good performance. To this end, we measure the quality of an iris segmentation thanks to a GMM model trained on good quality iris texture and we use a threshold on this quality value to shift between the 2 versions of OSIRIS. We show on ICE2005 database how the choice of this threshold value allows compromising between performance and processing speed of the complete process.