Fingerprints ha
ve been accepted as an unchangeable, unique identifier for human identities since the 1890’s and have been in use by law enforcement ever since. From that point on, fingerprint recognition has evolved and has been in operational use for decades. Most modern smartphones have fingerprint capture devices and many restricted areas are protected using fingerprints as a means of identification.
In recent years, especially with the spread of the COVID-19 pandemic, the need for more hygienic alternatives has gained awareness. One of those alternatives is the field of contactless fingerprint recognition. Though this field was thoroughly researched over the years, most feature extraction al-
gorithms are still designed for the contact-based domain, with the lack of sufficient contactless training data being one of the biggest challenges.
This thesis examines contactless fingerprint feature extraction. It will explore two main topics: First, FingerNet, an algorithm for detecting level two features in the contact-based domain is retrained with contactless data. Hence, we present and evaluate a retraining workflow. The retrained FingerNet model was evaluated on two real world databases and compared to the original model. Second, multiple Convolutional Neural Network (CNN)s are trained to classify fingerprints based on their fingerprint patterns. We conducted 37 exhaustive experiments for classifying level one features and
evaluated our trained models on three real world databases: PolyU, ISPFDv1,
and HDA.
While retraining FingerNet leads to comparable results as the original, we show that synthetic data can be used to adapt algorithms designed for the contact-based domain to the contactless domain. Additionally, we show that CNNs trained on synthetic, contactless data are a promising method to
classify fingerprint patterns, although further research is needed to improve results.