Cremer, SandraDorizzi, BernadetteGarcia-Salicetti, SoniaLempérière, NadègeBrömme, ArslanBusch, Christoph2018-11-192018-11-192012978-3-88579-290-1https://dl.gi.de/handle/20.500.12116/18321The most common iris recognition systems extract features from the iris after segmentation and normalization steps. In this paper, we propose a new strategy to select the regions of normalized iris images that will be used for feature extraction. It consists in sorting different sub-images of the normalized images according to a GMM-based local quality measure we have elaborated and selecting the N best sub-images for feature extraction. The proportion of the initial image that is kept for feature extraction has been set in order to compromise between minimizing the amount of noise taken into account for feature extraction and maximizing the amount of information available for matching. By proceeding this way, we privilege the regions for which our quality measure gives the highest values, namely regions of the iris that are highly textured and free from occlusion, and minimize the risks of extracting features in occluded regions to which our quality measure gives the lowest values. We also control the amount of information we use for matching by including, if necessary, regions that are given intermediate values by our quality measure and are free from occlusion but barely textured. Experiments were performed on three different databases: ND-IRIS- 0405, Casia-IrisV3-Interval and Casia-IrisV3-Twins, and show a significant improvement of recognition performance when using our strategy to select regions for feature extraction instead of using a binary segmentation mask and considering all unmasked regions equally.enText/Conference Paper1617-5468