Segmentation-level Fusion for Iris Recognition

Transcription

1 Segmentation-level Fusion for Iris Recognition Peter Wild 1,3, Heinz Hofbauer 2, James Ferryman 1, Andreas Uhl 2 1 School of Systems Engineering, University of Reading, Reading RG6 6AY, UK. 2 Dept. of Computer Sciences, University of Salzburg, 5020 Salzburg, Austria. 3 AIT Austrian Institute of Technology GmbH, 2444 Seibersdorf, Austria. peter.wild@ait.ac.at, {hhofbaue, uhl}@cosy.sbg.ac.at, j.m.ferryman@reading.ac.uk 14 th Int l Conf. of the Biometrics Special Interest Group (BIOSIG) September 2015 P. Wild et al.: Segmentation-level Fusion for Iris Recognition 1/20

2 Outline 1 Introduction 2 Multi-segmentation Fusion Methods 3 Experimental Study 4 Conclusion P. Wild et al.: Segmentation-level Fusion for Iris Recognition 2/20

3 Motivation Challenge Existing: fusion methods at data/feature, score and rank/decision level. Widely ignored: fusion at normalisation/segmentation level prior to feature extraction. Missing: any standards for interchange of segmentation results Ambition Motivation 1: better accuracy for less invasive recording conditions? Motivation 2: potentially faster alternative to multi-algorithm fusion? Motivation 3: improved understanding of types of segmentation errors. Impact Investigate and suggest methods for effective multi-segmentation fusion, tested on public datasets with open source software. P. Wild et al.: Segmentation-level Fusion for Iris Recognition 3/20

4 Related Work Super-resolution [Huang et al. BMVC 03] among first data-level fusion aproaches for iris present a Markov network learning-based fusion method to enhance the resolution of iris images Iris image-fusion [Hollingsw. et al. TIFS 09] combine high-res. images from multiple frames [Jillela et al. WACV 11] image-level fusion with Principal Comp. Transform [Llano et al. ICB 15] PCA vs. Laplacian Pyramid & Exp. Mean fusion Segmentation fusion [Uhl et al. ICIAR 13] proof-of-concept human (manual) ground truth segmentation: 97.46% % GAR at 0.01% FAR no automated algorithms P. Wild et al.: Segmentation-level Fusion for Iris Recognition 4/20

5 Fusion Framework Iris Segmentation I Segmentation Algorithm 1... {P1, L1, E U 1, E L 1 } N1 Fusion {P, L, E U, E L } Iris Texture Rubbersheet transform Segmentation Algorithm k {Pk, Lk, E U k, EL k } Nk N Noise Mask Input: inner/outer boundaries P, L : [0, 2π) [0, m] [0, n]. Output 1: refined boundaries for Rubbersheet mapping R(θ, r) := (1 r) P(θ) + r L(θ). Output 2: texture and noise masks T, M : [0, 2π) [0, 1] C (C is the target color space, M = N R, T = I R for the original n m image I and noise mask N). P. Wild et al.: Segmentation-level Fusion for Iris Recognition 5/20

6 Investigated Questions 1 Does the combination of automated iris segmentation results yield more accurate result than each of the employed original segmentation algorithms? 2 How does the choice of database and segmentation algorithms impact on iris segmentation fusion? 3 How do outliers impact on overall recognition accuracy and how do ground-truth-based vs. recognition-based evaluations relate to each other? Contribution analysis of reference methods for iris segmentation-level fusion; considering ground-truth and recognition-based assessment. P. Wild et al.: Segmentation-level Fusion for Iris Recognition 6/20

7 Error Measures Ground-truth evaluation: Assessing segmentation noise mask; measures suggested by Noisy Iris Challenge Evaluation - Part I (NICE.I) and F-measure [Hofbauer et al. ICPR 14]: E 1 := 1 k k i=1 fp i + fn i ; E 2 := 1 mn 2 ( 1 k F-measure = F 1 := 1 k k i=1 k i=1 fp i fp i + tn i ) ( 1 k tp i tp i (fn i + fp i ) k i=1 fn i ) fn i + tp i (1) Recognition-based evaluation: Account for feature-based tolerance of false segmentations. Use Equal Error Rate (EER) as main performance indicator; McNemar test [McNemar, Psy. 47]. (2) P. Wild et al.: Segmentation-level Fusion for Iris Recognition 7/20

8 Segmentation Fusion Sum-Rule Interpolation: This fusion rule combines boundary points B i (θ) of curves B 1, B 2,... B k : [0, 2π) [0, m] [0, n] into a single boundary B, for pupillary and limbic boundaries, in analogy to the sum rule. B(θ) := 1 k B i (θ); (3) k i=1 Augmented-Model Interpolation: This model combines boundaries B 1,..., B k within a jointly applied parametrisation model ModelFit minimizing the model-error (e.g., Fitzgibbon s ellipse-, or least-squares circular fitting), executed separately for inner and outer iris boundaries. Models are combined, not only points. ( k B(θ) := ModelFit B i )(θ) (4) i=1 P. Wild et al.: Segmentation-level Fusion for Iris Recognition 8/20

9 Iris Scanning and Pruning Process Overview over the iris scanning and pruning process vertical scana area horizontal scan area (a) With outliers µr + 2.5σr µr 2.5σr r Cr outlier outlier (b) With outliers pruned Input: Segmentation masks N; Method: Augmented model interpolation based on mask-scan; N equidistant scan lines are used to generate points; Outlier detection and removal using center of gravity C r (z-score > 2.5). P. Wild et al.: Segmentation-level Fusion for Iris Recognition 9/20

10 Iris Scanning and Pruning Process: Details High number of scan lines is desirable ( N = 100 ); If the mask contains holes (noise), they should be closed by an dilate+erode morphological operation; OSIRIS algorithm produces masks, which extend over the actual boundaries, therefore a restriction step is introduced. Actual mask is generated by fitting an ellipse to the point clouds by a least-squares method. (a) Original (b) Corrected boundaries (c) Without noise P. Wild et al.: Segmentation-level Fusion for Iris Recognition 10/20

11 Tested Segmentation Algorithms CAHT - Contrast Adaptive Hough Trans. [Rathgeb et al. 2012] traditional sequential (limbic-after-pupillary) method; based on circular HT and contrast-enhancement; WAHET - Weighted Adaptive HT & ET [Uhl et al. BTAS 12] two-stage adaptive multi-scale HT segmentation, elliptical; OSIRIS - Open Source for Iris [Petrovska et al. 2007] circular HT-based method with boundary refinement; IFPP - Iterat. Fourier Pulling & Pushing [Uhl et al. ICIAR 12] iterative Fourier-series approximation and pulling and pushing ; P. Wild et al.: Segmentation-level Fusion for Iris Recognition 11/20

12 Impact on Recognition Accuracy Equal error rate (EER) for combinations using USIT (Uni Salzburg Iris Toolkit [Rathgeb et al. 2012]) algorithms Ma (wavelet zero-crossing based) and Masek (1D Log-Gabor): Casia v4 Interval database Equal-error rate [%] of Masek CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP 8.10 IIT Delhi database Equal-error rate [%] of Masek CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP 3.87 Equal-error rate [%] of Ma CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP 8.78 Equal-error rate [%] of Ma CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP 4.36 P. Wild et al.: Segmentation-level Fusion for Iris Recognition 12/20

13 Results of the McNemar test, reported as X 2 values fused with fused with Casia v4 Interval database X 2 statistic for Masek single method CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP X 2 statistic for Ma single method CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP IIT Delhi database X 2 statistic for Masek single method CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP X 2 statistic for Ma single method CAHT WAHET OSIRIS IFPP CAHT WAHET OSIRIS IFPP P. Wild et al.: Segmentation-level Fusion for Iris Recognition 13/20

14 Results of Segmentation-level Fusion (Recognition) Segmentation fusion increased performance in 10 out of 24 combination scenarios; Only one setup, IFPP + WAHET, which consistently increases the performance; Only one case, OSIRIS + CAHT using Ma on IITD deteriorates performance of both individual results. McNemar tests using χ 2 approximation with the continuity correction proposed by Edwards reveals EERs are different (critical value X indicates a rejection of the null hypothesis with at least 99% significance). P. Wild et al.: Segmentation-level Fusion for Iris Recognition 14/20

15 Ground-truth Segmentation Accuracy Good vs bad segmentation-fusion Casia v4 Interval database Segmentation error [%] E 1 E 2 Good Bad Good Bad CAHT WAHET (NIR) Fusion (Sum Rule) IIT Delhi database Segmentation error [%] E 1 E 2 Good Bad Good Bad CAHT WAHET (NIR) Fusion (Sum Rule) Sum Rule segmentation fusion performance on good versus bad segmentations (distance of centers/radii); E 1, E 2 test: Fusion accuracy on the good set improved, while averaging performance for the bad set; F-Measure test: Fusion exhibits a closer conformity to the ground truth than each individual segmentation algorithm reduction in outliers. P. Wild et al.: Segmentation-level Fusion for Iris Recognition 15/20

16 F-Measure Test (Casia-v4-Interval Ground-truth) IFPP WAHET IFPP+WAHET S1249R02 S1228L06 S1212L07 S1196R03 S1180R05 S1167R08 S1159R03 S1144R03 S1135L07 S1125R05 S1116R01 S1107L05 S1098L01 S1089L02 S1079R05 S1069R02 S1057R04 S1042R03 S1028L07 S1011R02 S1001L Subject id F-measure F-measure F-measure P. Wild et al.: Segmentation-level Fusion for Iris Recognition 16/20

17 Possible Effects of Combining Masks Positive Neutral (a) Shape mismatch correction (b) Boundary mismatch correction due to cut-off iris (c) Discrepancy Negative (d) Matching errors (a) Detection flaw (b) Missed boundary (c) Pruning failure P. Wild et al.: Segmentation-level Fusion for Iris Recognition 17/20

18 Conclusion Investigated topic Multisegmentation fusion using pairwise combinations of CAHT, WAHET, IFPP and OSIRIS iris segmentation algorithms. Results 10/24 cases: autocorrective behaviour (augmented model fusion); best: 0.64% EER for WAHET+CAHT vs 0.99% EER for CAHT; better correction, if iris undershoots rather than overshoots; non-convex and miss shaped masks can lead to fusion problems; ground-truth evaluations miss corrective behaviour for outliers. Next steps advanced, sequential approaches with more than 2 algorithms taking processing time into account. P. Wild et al.: Segmentation-level Fusion for Iris Recognition 18/20

19 References F. Alonso-Fernandez and J. Bigun. Quality factors affecting iris segmentation and matching. In Proc. Int l Conf. on Biometrics (ICB), H. Hofbauer, F. Alonso-Fernandez, P. Wild, J. Bigun, and A. Uhl. A Ground Truth for Iris Segmentation. In Proc. 22nd Int l Conf. Pattern Rec. (ICPR), E. Llano, J. Vargas, M. Garca-Vzquez, L. Fuentes, and A. Ramrez-Acosta. Cross-sensor iris verification applying robust fused segmentation algorithms. In Proc. Int l Conf. on Biometrics (ICB), 2015, pages 1 6, Q. McNemar. Note on the sampling error of difference betw. correlated proportions of percent. Psychometrika, 12(2): , C. Rathgeb, A. Uhl, and P. Wild. Iris Recognition: From Segmentation to Template Security, vol. 59 Adv. Inf. Sec.. Springer, P. Wild et al.: Segmentation-level Fusion for Iris Recognition 19/20

20 Thank you for your attention! Any questions? The work has been supported by the FastPass project. The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/ ) under grant agreement no This publication only reflects the authors view and the European Union is not liable for any use that may be made of the information contained therein. All document contained therein cannot be copied, reproduced or modified in the whole or in the part for any purpose without P. Wild et al.: Segmentation-level Fusion for Iris Recognition 20/20