Page - 147 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Anguli Verifinger x, y, t U -Shaped NeuralNetwork Dice Loss Ground Truth Ridge Pattern Noised synthetic fingerprints Minutiae list Groundtruth minutiae blocks Output minutiae blocks Fig. 2. The whole processing pipeline used for training the minutiae extraction network. TheU-shapednetworkarchitectureappliedhere isusedfor medical segmentation applications [19], [7]. Training deep neural networks is also the focus of [9], where residual connections are used to allow training of very deep neural networks. Research into making residual connection better is reported in [12], [27], [7]. The GAN framework is first introduced in [8]. Improve- ments to the stability of adversarial training are proposed in [20], [3]. Based on the results in GANs a refinement network is introduced in [21] for the gaze direction of eye images. In our work a similar approach is used to refine fingerprints. III. MINUTIAE EXTRACTION USING CNN The ground truth minutiae list is turned into a binary image by creating an image with the same shape as the corresponding fingerprint and setting every pixel to zero. Theneverypoint inaminutiae region is set toone.Aminutia region is defined as a 7×7 pixel square encapsulating the minutia landmark as its centroid. Our deep neural network is used to find a mapping from the input fingerprint to this binary image. This procedure turns the task into a binary segmentation problem. A. Training Pipeline The synthetic fingerprint generator Anguli [2] is used to generate a training set. As can be seen in Fig. 2 Verifinger is used to extract the ground truth of the original ridge pattern. For the purpose of this algorithm it is assumed that the minutiae extractor works perfectly on the ridge pattern. Therefore the estimated bifurcations and terminations of the ridge image in Fig. 2 are input to the learning stage as well as ground truth for evaluation. The deviation of the minutiae map and the network output is calculated using dice loss (1), whereα is a smoothing factor. Dice loss is reported to produce almost binary outputs [7]. loss=− 2ypredytrue+α ∑ypred+∑ytrue+α (1) B. Network Architecture The base architecture of the models used in this work can be seen in Fig. 3 and builds on the U-Shaped Network pioneered in [19]. The key differences are: 1) Strided convolution instead of pooling to learn down- sampling filters. 2) 224×224 crop to preserve the aspect ratio of the fingerprints. 3) Layer blocks on intermediate levels of the U-Shaped Network instead of pure convolutions. 4) Batch Normalization [11] before every convolution. 5) Dropout with a probability of 0.5 before the final Convolution Layer. 6) Upsampling is done by repeating the pixel in a 2×2 window. Then the upsampling feature maps are con- catenated with the output feature maps of the layer block on the same level in the downsampling path. Finally batch normalization, a Regularized Linear Unit (ReLU) activation function and a 3×3 convolution are applied to all the feature maps, before they are passed on to the next layer block. To preserve information flow, the amount of filters is doubled, when the size of the input data is reduced, as observed in [22]. The layer blocks on specific levels vary in the number of filters used. A model is build with only Wide Residual Blocks [27] (WRN), one with only Densely Connected Blocks [12] (DenseNet) and one with only Bot- tleneck Residual Blocks [7] (ResUnet). In total, each model used in this work has approximately 8 million parameters. C. Extracting a Minutiae List The output of the neural network is a binary minutia regions map. For biometric authentication, a list of minutia points with quality and orientation is needed. For the final position of the minutiae the connected components of the binary map are used. The centroid of each component represents one minutia position. The area a of the connected 147