Introduction to the Special Issue on the Recognition of Visible Wavelength Iris Images Captured At-a-distance and On-the-move

This special issue regards the recognition of degraded iris images acquired in visible wavelengths. During 2009 and 2010, the University of Beira Interior (Portugal) promoted two International evaluation initiatives about this subject, named Noisy Iris Challenge Evaluation (NICE) I and II. The first one focussed on the evaluation of iris segmentation strategies, considering that iris data acquired in visible wavelengths (VW) usually has much higher level of detail than traditionally used near infra-red data (NIR), but also has many more noise artifacts, including specular and diffuse reflections and shadows. Also, the spectral reflectance of the sclera is significantly higher in the VW than in the NIR and the spectral radiance of the iris with respect to the levels of its pigmentation varies much more significantly in the VW than in the NIR. The NICE:II contest complemented its predecessor in terms of the traditional pattern recognition stages, evaluating different signature encoding and matching strategies. In order to guarantee that unbiased performance measures were obtained, all the participants used the exact same segmented data, which were automatically obtained according to the highest performing method in the NICE:I. Again, participation in NICE:II was free of charge and opened to all research and academic institutions. Sixty-seven participants from thirty countries registered in the contest 1 and received a training set composed of 1000 images and the corresponding binary iris segmentation masks. The task assigned to participants is illustrated in Fig. 1: to construct a binary executable that receives (by command-line parameters) a pair of iris samples and their iris segmentationmasks and outputs a text file containing a score that corresponds to the dissimilarity between the irises. This score d should be a metric, i.e., it should meet the following conditions: (1) d(I, I) = 0; (2) d(I1, I2) = 0) I1 = I2 and (3) d(I1, I2) + d(I2, I3) P d(I1, I3). In the evaluation, disjoint sets of 1000 unseen images and the corresponding segmentation masks were used to rank participants. Let I 1⁄4 fI1; . . . ; Ing be a set of iris images, M 1⁄4 fM1; . . . ;Mng their binary iris segmentation masks and id(.) the identity function on an image. An one-against-all comparison scheme yields a set of match D 1⁄4 di1; . . . ; d i m n o and of non-match D 1⁄4 de1; . . . ; d e k dissimilarity scores, respectively, for the cases where id(Ii) = id(Ij) and id(Ii)– id(Ij). As suggested by Daugman, for two-choice decisions (e.g., match/ non-match) the decidability index d0 measures how well separated the two types of distributions are, and recognition errors correspond to their overlap area: