Seite - 149 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

(a) Linear Transform Augmentation (b) Gaussian Augmen- tation (c) Morphological Augmentation (d) Distorted Ground Truth Ridge Pattern (e) Refinement Net- work Input (f) Refinement Network Output with minutiae regions. Fig. 6. Illustration of the refined data. resized image is used as input to the network. An example input and output image can be seen in Fig. 6e and Fig. 6f. B. Refinement Network The Refinement Network used in this work is based on the GANparadigm,whereadualoptimizationproblemissolved. A refiner and a discriminator network are simultaneously trained against each other. The refiner network tries to fool the discriminator by applying refinements to a synthetic fingerprint, while the discriminator is used to discriminate between fake and real data. The purpose of such a network is to find a Nash Equilibrium [20] where both networks are optimal. The only application of a refinement network to our knowledge is in [21]. In our work the approach therein is extended by using noise on the input data to improve the stability of training such a network [3]. One key observation is that using the input image itself for regularization is limiting the amount of possible refinements for fingerprints. Here we propose to use the Hessian of the image instead of the image itself for regularization. The Hessian represents the actual ridge pattern of the fingerprint independent of the pixel intensity values. Mean Squared Error is used to penalize deviation from the Hessian, while the refiner network still needs to fool the discriminator network. The refiner network uses the same architecture as the minutiae extraction network (Fig. 3), only smaller in the number of layer blocks and filters. Wide residual blocks [27] are used for every layer block starting with 32 filters and doubled on its way down and halved on their way up. Fingerprints like in Fig. 6f are produced by this method. Here, the problem observed by current synthetic fingerprint refiners of modeling noise is addressed by using such a network [5]. V. EXPERIMENTS This section showcases the results obtained with our method. All our models were programmed using the python framework Keras [6] and trained on a Nvidia Geforce Titan X. For training, the Adam [13] optimizer is used with an initial learning rate of 0.001. The learning rate is cut in half, if the validation error has not decreased for three consecutive epochs. For other minutiae extraction algorithms, an Intel Xeon - W3550 CPU was used. A. Experimental Setup For training 28.000 fingerprints with five impressions per fingerprint were generated using Anguli. In total 140.000 fingerprints were used for training, which included a vali- dation set of 10.000 fingerprints. The different impressions can be seen in Fig. 1. Out of the impressions three contain medium noise and the other two use little and heavy noise respectively. Non linear distortions are used on 3.000 of those fin- gerprints and on all of their impressions. All the other augmentations, as described in Section IV are applied on the fly. An annotated real dataset of 300 fingerprints constructed from 220 samples of the sd04 [26] and the 80 images of the fvc2000 DB4 B [15] dataset are used additionally to increase the effectiveness of the classifier. The real dataset used for the refinement network is the UareU [1] dataset. B. Deep Learning Experiments In Fig. 9 the difference in performance for the various layer blocks defined in Section III can be seen. In contrast to thefindings in [12]usingdenselyconnectedblocksdidnot work as well for the minutiae detection problem. Bottleneck residual blocks performed similarly to wide residual blocks, which is similar to the findings in the original paper [27]. C. Experiments on FVC2000 databases Here, the performance of our method is compared to other minutiae extraction algorithms on the FVC2000 [15] dataset consisting of real world fingerprints. To match the minutiae against each other, the minutiae matcher BOZORTH [25] was used. The results of this experiment can be seen in Fig. 7 and Fig. 8, where GAR and FAR denote the Genuine Acceptance Rate and False Acceptance Rate accordingly. Using those metrics the Equal Error Rate (EER) can be calculated by finding the rate where (2) holds. GAR=1−FAR (2) The extracted EER of the evaluated minutiae extractors is shown in Table I. Our algorithm performs better than 149