Seite - 150 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

80 82 84 86 88 90 92 94 96 98 100 0.1 1 10 100 FAR [%] U-net FCNN MINDTCT Verifinger (a) Using only Anguli generated data. 80 82 84 86 88 90 92 94 96 98 100 0.1 1 10 100 FAR [%] U-net FCNN MINDTCT Verifinger (b) Using Augmentations 80 82 84 86 88 90 92 94 96 98 100 0.1 1 10 100 FAR [%] U-net FCNN MINDTCT Verifinger (c) Using Augmentations and Refinements. Fig. 7. Equal Error Rate Comparison on FVC2000 [15] DB 1 using synthetic, augmentated or refined data. 80 82 84 86 88 90 92 94 96 98 100 0.1 1 10 100 FAR [%] U-net FCNN MINDTCT Verifinger (a) Using only Anguli generated data. 80 82 84 86 88 90 92 94 96 98 100 0.1 1 10 100 FAR [%] U-net FCNN MINDTCT Verifinger (b) Using Augmentations 80 82 84 86 88 90 92 94 96 98 100 0.1 1 10 100 FAR [%] U-net FCNN MINDTCT Verifinger (c) Using Augmentations and Refinements. Fig. 8. Equal Error Rate Comparison on FVC2000 [15] DB 3 using synthetic, augmentated or refined data. Fig. 9. Model comparision between Dense Blocks, Wide Residual Blocks and Bottleneck Residual Blocks. MINDTCT on real datasets. Also we report clear perfor- mance improvements by using a refinement network. Addi- tionally it is also the fastest method, when run on a GPU. D. Sample Results for Refinement Network The only quality metric to our knowledge for GANs is the inception score [20], which is not applicable for our use case. Therefore, this section shows the visual result of the refinement network. In Fig. 10 we can see a comparison of using self regularized MSE versus the Hessian regularized version of the network. In the Hessian regularized examples the ridge pattern is better preserved and less artifacts are introduced into the refined fingerprint. E. Sample Results for Various Fingers Anillustrationonwhichminutiaeare foundusingdifferent training data is given in Fig. 11. Here, by training solely on TABLE I EQUAL ERROR RATE AND ENROLLMENT SPEED FOR FVC2000 [15] DATABASES Algorithm DB 1 DB 3 Time in sec. Synth. Unet FCNN 21.80% 32.75% 0.12 on gpu Augm. Unet FCNN 7.01% 16.63% 0.12 on gpu Ref. Unet FCNN 5.99% 9.42% 0.12 on gpu MINDTCT [25] 6.63% 12.11% 0.14 on cpu Verifinger [24] 3.28% 6.31% 1.08 on cpu (a) Self regularized GAN Output (b) Hessian regular- ized GAN Output. (c) Hessian regular- ized GAN Output. Fig. 10. Sample refiner network output images for self regularized and Hessian regularized training based on the GAN approach. synthetic fingerprints the minutiae map is clearly wrong as shown in Fig. 11. The network trained on augmented data outputs a subset of the correct minutiae. In contrast, the network trained on GAN data outputs a reasonable minutiae map for this example. An example of a clear mismatch between two images of the same fingerprint can be seen in Fig. 12. Even though the matching score is 0%, overlapping minutiae are found. However, the orientation does not match because of the noise in the fingerprint. 150