Seite - 148 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Layer BlockC Layer BlockD ∪ Shape: 112x112 Shape: 56x56 Shape: 28x28 Upsample A Upsample B ∪ ∪ LayerBlockB Sigmoid 1x1 Convolution Transition A Transition B Layer BlockC Shape: 224x224 Layer BlockA Upsample C Layer BlockB Layer BlockE Transition C Fig. 3. U-Shaped network architecture used for minutiae extraction. (a) (b) Fig. 4. Estimation of the orientation field for a sample fingerprint taken from the FVC2000 DB 1. components is used to determine the quality of the minutiae between 0−100 using quality=min(a∗2,100). The orientation of the minutiae is extracted using an orientation field as described in [10]. The orientation is estimated for every 16×16 region as visualized in Fig. 4b. IV. FINGERPRINT REFINEMENT The synthetic fingerprint generator Anguli [2] is used to generate random ridge patterns (Fig. 5a). Then multiple variations of every ridge pattern are generated by using different noise models as can be seen in Fig. 5(b,c). Each variation is called an impression of that particular ridge pattern. Because Anguli does not output the minutiae in- formation, a commercial minutiae extractor, Verifinger [24], is used to extract the minutiae data out of the ridge pattern. For the purpose of this paper it is assumed that Verifinger works perfectly on the binary ridge pattern. Those minutiae landmarks are then used for all the impressions (Fig. 5(a-c)). A. Augmentation on Synthetic Fingerprints By comparing Fig. 1(d-f) with Fig. 5(a-c) the differences between real and synthetic fingerprints are easily spotted. To bridge this gap the following augmentations are used: 1) Non linear distortions: Tomodel the contact regionof a fingerprint, random non-linear distortions are used. (a) (b) (c) Fig. 5. Anguli [2] generated ridge pattern with two different impressions and the minutiae extracted using Verifinger [24] . This also introduces changes in local ridge frequency to synthetic fingerprints as can be seen in Fig. 6d. The distorted ridge pattern is used by Anguli to generate new impressions. 2) Morphological operations: Grayscale Dilation and Erosion are used to model wet and dry fingerprint images [5]. An example of this can be seen in Fig. 6c. 3) Random rotation, translation and shearing: Fin- gerprint images are randomly translated, rotated and sheared to gain invariance to linear transformations. An example of this can be seen in Fig. 6a. 4) Random Blurs: The images are randomly blurred with a Gaussian kernel, where the variance varies to simulate noisy fingerprints as can be seen in Fig. 6b. 5) Random Mirroring: Fingerprint images are randomly mirrored either horizontally or vertically with a 0.5 probability for each direction. 6) Refinement Network: A Refinement Neural Network, based on GANs is used to refine images to look more like real world fingerprints. The input size to the network is 224×224. Therefore synthetic fingerprints are resized by a random factor between zero and the difference in image dimension, while keeping the aspect ratio. Then a random 224×224 crop of the 148