Slovak University of Technology, Bratislava, Slovak Republic

Faculty of Electrical Engineering and Information Technology

Computer and Network Security

The primary goal of this thesis is to design and develop a complex fingerprint recognition system, referred to as DBOX-server. It runs in Linux environment, providing an efficient way of identity verification in computer network. The most important thesis goals are image preprocessing, fingerprint feature extraction using convolutional neural network and subsequent feature matching. In order to make mass fingerprint processing feasible, we employed a GPU computing power along with CPU threads. In the introductory part, we present our goals and describe all frameworks we used during the software implementation. The following sections of the thesis deals with design of our solution, organising individual tasks into standalone libraries. At this stage, deep neural networks play a crucial role, as we employed them as minutiae classifiers. Moreover, we created a specialized tool for image data collection, training and interactive testing of neural networks. For this purpose we used Caffe library. We also created our own multithreaded matching algorithm based on well known Bozorth3 method developed by NIST. In addition, during the testing stage we also made use of Suprema matching technology. In the final section we discuss the obtained results of comprehensive tests carried out on popular fingerprint databases. Our solution was evaluated in terms of speed and recognition rate. Finally, we compared the results with published works in the field and proposed future improvements.