Deep Learning for Fingerprint Recognition Systems
Patrick Schuch · NORA - Norwegian Open Research Archives · 2019
Summary Biometric recognition is a typical means to identify individuals or to verify claimed identities. Use cases are manifold. For example, users can unlock their smartphones for convenience by presenting their faces or fingerprints. Or one’s identity is verified when crossing borders. Today, biometric recognition already has many points of contact with our daily life and there are more to come. Besides iris and face, fingerprint is the most wide spread biometric trait used for recognition. Fingerprints are assumed to be unique for each and every finger. This makes it an ideal trait for recognition. In addition, fingerprint recognition has more than a century of tradition in the field of biometric recognition. A great amount of expertise and engineering skill made it a quite mature technology over time. Only few false positive and false negative errors are made in recognition in today’s deployed systems. However, fingerprint recognition is still far from being perfect. In contrast to popular opinion, fingerprint recognition is not a solved problem. Actually, there is still a lot of work to do. As biometric systems become larger and become more inclusive, even new challenges arise. Systems need to deal with large amounts of data while keeping performance with respect to recognition performance as well as transaction times in a reasonable order. Recognition shall work for everyone and shall not exclude a certain ethnic group or subset of the population. It will work in unconstrained conditions. However, it shall still make no erroneous decisions. Engineering may have come to its limits at this stage. In contrast to classical engineering, machine learning based on artificial neural networks may be a reasonable alternative. The emerging technologies of Deep Learning achieve tremendous successes in many domains of image processing and pattern recognition. This work assesses the application of such innovative machine learning concepts to fingerprint recognition. Three central aspects and challenges in fingerprint recognition are inspected in detail: fingerprint sample enhancement, orientation field estimation, and efficient processing structures.