Hatim Dali defended his Master Thesis on: a video-based approach on contactless fingerprints
Fingerprints are among the most widely used biometric features and are an integral part of today’s life. Since the technologies used have been developed steadily over the last few decades, they are now also used as a safety measure in ordinary smartphones. A promising alternative to traditional detection systems, which may have problems with latent fingerprints or uneven contact pressure, is in contactless systems.
The aim of this work is to develop and evaluate an Android app for contactless, video-based capturing of fingerprints. Hereby, up to four fingers are to be recorded and processed simultaneously. Several steps are necessary to extract the fingerprints from the images, which are obtained from the incoming video stream. These include detecting the fingers and calculating a correction angle to obtain a uniform orientation. Moreover, a separation of the individual finger is needed. In order to ensure a high quality of fingerprints, algorithms are also implemented which sort out unusable images at an early stage. Lastly, the fingerprints on the images thus obtained are pro grammatically enhanced. The first part of the work consists of an overview on current approaches to the individual steps required for the app. Afterwards the developed pipeline is explained in detail, as well as the criteria of the test database is called.
In the absence of a publicly accessible database with the required data records, this has to be created by oneself. In addition to the finger images, metadata about the test subjects are stored here. These include, amongst others, the age, sex, and skin color of the subjects. Furthermore, image-specific information is recorded, such as the resolution of the image and the correction angle. Several performance measurements were defined for the evaluation of the developed app. The first step is to evaluate how much of the data can be used for further tests, and what relationships exist between the errors that have occurred. Subsequently, the Equal Error Rate ( EER ) of each color space of the two smartphones in this thesis will be evaluated, as well as the best performing color space for both smartphones.
For the final evaluation of the developed system 1620 test images were collected. A first manual control showed that about 40% of the data is faulty. Only a quarter of this shortcomings could be discovered through a programmatic approach. In about 80 of the pictures no minutiae could be detected although these were classified as valid during the manual control. The usable data yielded an EER of about 26,6% and 24,4% for enhanced images on both test devices. Grayscale, as well as RGB images performed worse with about 33.4% to 41,7%. As soon as the same tests were repeated with the previously sorted data, the EER increased significantly in all test scenarios.