Technical Report

Cross-sensor comparison experimental results reported here show that the procedure defined and simulated during the Cross-Sensor Comparison Competition 2013 by our team for migrating / upgrading LG2200 based to LG4000 based biometric systems leads to better LG4000-to-LG2200 cross-sensor iris recognition results than previously reported, both in terms of user comfort and in terms of system safety. On the other hand, LG2200-to-LG400 migration/upgrade procedure defined and implemented by us is applicable to solve interoperability issues between LG2200 based and LG4000 based systems, but also to other pairs of systems having the same shift in the quality of acquired images. 10 http://lmrec.org/acstl/tr-460.pdf 11 http://lmrec.org/acstl/cross-sensor-competition-team-2013/ 12 http://lmrec.org/acstl/ 13 http://lmrec.org/bodorin/ 14 http://www3.nd.edu/~asgroi/Competition/CrossSensorCompetition.htm 15 http://engineering.nd.edu/profiles/asgroi 16 http://www3.nd.edu/~kwb/ 17 http://www3.nd.edu/~flynn/ 18 http://www3.nd.edu/~cvrl/ Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison 4 I. Iris Segmentation, Encoding and Matching Procedures for LG4000-to-LG2200 Cross-Sensor Iris Recognition The methods used by us to process the data sets given in the Cross-Sensor Comparison Competition 2013 comply to the ISO/IEC 19794-6 [5] data interchange format standard (Fig. 1). The eye images of KIND 1 are processed in order to detect the pupil center and the pupil boundary. Using this information, a region of interest is selected from the initial eye image and represented as an image of KIND 48, which is further unwrapped as a KIND 16 unsegmented pupil-centered polar image (Fig. 2). The limbic boundary is found in the unwrapped version of the region of interest (KIND 16) by applying CFIS2 (Circular Fuzzy Iris Segmentation, second variant, [1]). As a result, the unwrapped iris segment is obtained (Fig. 3). The segmented iris image (Fig. 3) is further encoded as a binary iris code (Fig. 4) using the Log-Gabor encoder, as in [1]. Hamming similarity (HS) expresses the matching between two iris codes: HS = ‖codeA == codeB‖ Length(codeA) (1) In the scenario of migrating/upgrading a LG2200 based system to a LG4000 based system, we consider that the LG2200 iris codes are simultaneously available for comparison with LG4000 iris code candidates. This allows us to assemble LG2200 binary iris codes representing the same eye as an LG2200 digital identity [2] and to compute – for any LG4000 iris code a fuzzy degree of membership to these LG2200 digital identities. Segmentation, digital identity encoding and matching procedures used for obtaining the results presented here are all proprietary. LG2200-toLG400 migration/upgrade procedure defined and implemented by us is applicable to solve interoperability issues between LG2200 based and LG4000 based systems, but also to other pairs of systems having the same shift in the quality of acquired images. Fig. 1: ISO/IEC 19794-6 data interchange format standard N. Popescu-Bodorin et al. 5 Fig. 2: Unwrapped region of interest (KIND 16) Fig. 3: Segmented polar image Fig. 4. Iris code of size 32x360 The following sections present the results obtained for LG4000 and LG2200 small, medium and large datasets, respectively. Each section contains figures illustrating the imposter / genuine score distributions and FAR / FRR curves, in linear and logarithmic scales. The figures 5, 6, 10, 11, 15, 16 and Table 1 illustrate the separation of the genuine and imposter score distributions (obtained for small, medium and large databases) in terms of their means and their standard deviations. Comparison Distribution Mean Standard Deviation Decidability Small LG4000-to-LG2200 cross-sensor recognition: Imposter 0.49845 0.0136 5.5779 Genuine 0.66322 0.0395 Medium LG4000-to-LG2200 cross-sensor recognition: Imposter 0.49874 0.0127 5.6386 Genuine 0.66038 0.0385 Large LG4000-to-LG2200 cross-sensor recognition: Imposter 0.49882 0.0124 5.7255 Genuine 0.66065 0.0380 Table 1: Means and Standard Deviations For each comparison test (small, medium and large), a decidability index is computed accordingly to the formula (2), d = |μ1−μ2| 1 2 (σ1 2 + σ2 ) (2) where: d is the decidability index; μ1 is the mean of the imposter scores; μ2 is the mean of the genuine scores; σ1 is the standard deviation of the imposter scores; σ2 is the standard deviation of the genuine scores; The separation between the genuine and imposter score distributions is also illustrated in Table 2 and figures 6, 11, 16 in terms of minimum genuine similarity score and maximum imposter similarity score. The overlapping between the imposter and genuine score distributions can be viewed as a performance criterion of the recognition technique, but also as a quality measure for the database of eye images. It must be noted that in Table 2 and also in the entire present report, is the Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison 6 SigSets 2013 specification that defines the imposter and the genuine comparisons, not otherwise. However, as the Figures 6, 11, and 16 show, there are some inconsistencies in the way in which the SigSets 2013 specification defines imposter and genuine comparisons: there are indeed some eye images enrolled under wrong IDs our visual inspection of the databases confirmed that. These cases lead up to the fact that some genuine and some imposter comparison defined by the SigSets 2013 are, in fact, mislabeled. Therefore, some so-called imposter scores defined by the SigSets are actually true genuine scores and some so-called genuine scores defined by the SigSets are actually true imposter scores indeed. Inevitably, this situation alters the FAR and FRR curves, and also the EER points as thresholds (abscises) and values (ordinates) making them to be slightly pessimistic estimates of the actual ones. The existence of the mislabeled eye images within the Cross-Sensor Comparison Database (and consequently the existence of the corresponding mislabeled imposter and genuine comparisons within the SigSets specification) is publicly affirmed here for the first time, despite the fact that the databases are in use for quite a while now. The detection of these cases of mislabeled eye images validates our work and our iris recognition technique. Naturally, a good iris recognition technique must be able to detect mislabeled eye images, especially when the number of eye images within databases is large – which is the case here, in Cross-Sensor Comparison Competition 2013. A clean-up of the Cross-Sensor Comparison Database and the corresponding update of the SigSets 2013 specification are both necessary. We will approach this task in our future work. Database Minimum genuine score Maximum imposter score Small 0.4605 0.6747 Medium 0.4485 0.6797 Large 0.4715 0.6947 Table 2: Minimum genuine scores and maximum imposter scores Database Figure FAR, FRR, threshold Small Fig. 9 (1E-3, 0.009267, 0.5506), (1E-4, 0.01655, 0.5636), (1E-5, 0.02863, 0.5776), (1E-6, 0.1021, 0.6117) Medium Fig. 14 (1E-3, 0.00941,0.5476), (1E-4, 0.01512, 0.5606), (1E-5, 0.02377, 0.5736), (1E-6, 0.06829, 0.6006) Large Fig. 19 (1E-3, 0.009723, 0.5476), (1E-4, 0.0144, 0.5596), (1E-5, 0.02361, 0.5726), (1E-6, 0.138, 0.6207) Table 3: Triplets of FAR, FRR, and threshold values detected for small, medium and large LG4000-to-LG2200 comparisons Database Figures EER value @ EER threshold Small Fig. 7, Fig. 8 0.005669 @ 0.5386 Medium Fig. 12, Fig. 13 0.006733 @ 0.5356 Large Fig. 17, Fig. 18 0.006167 @ 0.5356 Table 4: EER values detected for small, medium and large LG4000-to-LG2200 comparisons Based on the FAR/FRR curves presented in the figures 7, 8, 12, 13, 17 and 18, the EER points were located at the following thresholds: 0.5386 (Fig. 8), 0.5356 (Fig. 13), and 0.5356 (Fig. 18), with their corresponding values of approximately 5.66E -3, 6.73E -3, and 6.16E -3. For convenience, these data are also collected in Table 4. Figures 9, 14 and 19 show several triplets (FAR, FRR, threshold) describing four ways of balancing system security (expressed as FAR) and user discomfort (expressed as FRR). These data are also collected in Table 3. N. Popescu-Bodorin et al. 7 II. LG4000-TO-LG2200 IRIS RECOGNITION RESULTS OBTAINED ON SMALL LG4000 AND LG2200 DATABASES Fig.5: Imposter (μ=0.49845 σ=0.0136) and Genuine (μ=0.66322 σ=0.0395) Score Distributions (LinearY Scale) Fig.6 Imposter (μ=0.49845 σ=0.0136) and Genuine (μ=0.66322 σ=0.0395) Score Distributions (LogY Scale) Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison 8 Fig.7: False Accept curve, False Reject curve and EER point (LinearY Scale) Fig.8: False Accept curve, False Reject curve, and EER point (LogY Scale) N. Popescu-Bodorin et al. 9 Fig.9: Details on the False Accept curve and False Reject curve (LogY Scale) Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison 10 III. LG4000-TO-LG2200 IRIS RECOGNITION RESULTS OBTAINED ON MEDIUM LG4000 AND LG2200 DATABASES Fig.10: Imposter (μ=0.49874 σ=0.0127) and Genuine (μ=0.66038 σ=0.0385) Score Distributions (LinearY Scale) Fig.11: Imposter (μ=0.49874 σ=0.0127) and Genuine (μ=0.66038 σ=0.0385) N. Popescu-Bodorin et al. 11 Score Distributions (LogY Scale) Fig.12: False Accept curve, False Reject curve, and EER point (LinearY Scale) Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison 12 Fig.13: False Accept curve, False Reject curve, and EER point (LogY Scale) Fig.14: Details on the False Accept curve and False Reject curve (LogY Scale) N. Popescu-Bodorin et al. 13 IV. LG4000-TO-LG2200 IRIS RECOGNITION RESULTS OBTAINED ON LARGE LG4000 AND LG2200 DATABASES Fig.15: Imposter (μ=0.49882 σ=0.0124) and Genuine (μ=0.66065 σ=0.0380) Score Distributions (LinearY Scale) Fig.16: Imposter (μ=0.49882 σ=0.0124) and Genuine (μ=0.66065 σ=0.0380) Score Distributions (LogY Scale) Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison 14 Fig.17: False Accept curve, False Reject curve, and EER point (LinearY Scale) Fig.18: False Accept curve, False Reject curve, and EER point (LogY Scale) N. Popescu-Bodorin et al. 15 Fig.19: Details on the False Accept curve and False Reject curve (LogY Scale) Technical Report 460: Cross-Sensor Iris Recognition LG4000-to-LG2200 Comparison