Seite - 146 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Using a U-Shaped Neural Network for minutiae extraction trained from refined, synthetic fingerprints Thomas Pinetz1, Daniel Soukup 1, Reinhold Huber-Mo¨rk 1 and Robert Sablatnig2 Abstract—Minutiae extraction is an important step for robust fingerprint identification. However, existing minutia extraction algorithms rely on time consuming and fragile image enhancement steps in order to work robustly. We propose a new approach, combining enhancement and extraction into a Convolutional Neural Network (CNN). This network is trained from scratch using synthetic fingerprints. To bridge the gap between synthetic and real fingerprints, refinements are used. Here, an approach based on Generative Adversarial Networks (GANs) is used to generate fingerprints suited for training such a network and improving its matching score on real fingerprints. I. INTRODUCTION Because of their uniqueness and their temporal stabil- ity [10], fingerprint minutiae are a reliable way to determine the identityofan individual.Minutiaepointsare irregularities in ridge patterns, described using coordinates and orienta- tion [17]. Over 150 different irregularities in fingerprints have been identified [18]. While the amount of minutiae on a single fingerprint varies from finger to finger, there are approximately one hundred of such points comprising a regular fingerprint [17]. It was reported that only 10 - 15 minutiae are required to reliably identify an individual [17]. Currently fingerprint matchers like BOZORTH [25] work using minutiae landmarks. Extraction of minutiae is a hard problem though, which heavily relies on good quality fin- gerprint images [10]. To combat this, image enhancement algorithms are used [4]. Still, reliable minutiae extraction on arbitrary fingerprint images is an open problem as existing feature extractors largely rely on image quality (focus, reso- lution, skin condition, etc.) [23]. With the rise of deep learning in similar fields [7], [14], [19] and the availability of synthetic fingerprint generators [2], [5], it looks promising to use such methods for minu- tiae extraction. This paper contributes a new network for minutiae extraction following the idea to solve an equivalent segmentation problem. In this work the synthetic fingerprint generator Anguli [2] is used because of its availability. Anguli generates the training data needed as is shown in Fig. 1a. Because of the difference to real data as visualized in Fig. 1(d-f), augmentations are used (Fig. 1b) as described in Section IV. Here we contribute a novel technique to refine fingerprints based on the GANs [8] paradigm. An example output can be seen in Fig. 1c. Regularization is used to force the refinement network to retain the annotation data 1Austrian Institute of Technology, Donau-City-Straße 1, 1220 Wien {thomas.pinetz.fl, daniel.soukup, reinhold.huber-moerk}@ait.ac.at 2TU Wien, Karlsplatz 1, 1040 Wiensab@caa.tuwien.ac.at (a) Anguli generated fingeprint. (b) Augmented finger- print. (c) Refined fingerprint using our GAN based approach. (d) Finger taken from FVC2000 DB1. (e) Finger taken from FVC2000 DB3. (f) Finger taken from UareU dataset. Fig. 1. Illustration of the fingerprint data used in this work. (a-c are synthetic fingerprints, while d-f are real fingerprints.) while outputting a refined representation of the simulated fingerprint. The rest of the paper is organized as follows. Section II reviews related work. In Section III and IV the minutiae extractionalgorithmand the refinementmethodaredescribed in detail. In Section V the results obtained with our method arepresented.Finally inSectionVIwedrawourconclusions. II. RELATED WORK Minutiae detection for a sufficiently enhanced image is done by binarization of the grayscale image [10]. Currently fingerprint minutiae extractors use image enhancement rou- tines to achieve the desired quality [10], [4], [25], [24]. Recently there has been a similar approach to the minutiae extraction problem using a pre-trained Convolutional Neural Network, in a forensic setting [23]. However the CNN in [23] is used as a pre-processing step to find large regions containing a minutiae point. Then logistic regression and region pooling are used to extract the actual minutia position. In our approach the minutia extraction problem is rede- fined as a binary segmentation task, which the CNN solves directly. With our method there is no need for any time consuming pre- or post-processing. Additionally, synthetic fingerprint generators are used to train the network from scratch and make it suitable for the minutiae extraction problem. 146