Fingerprints are an extremely reliable method of identifying individuals in forensic science, as they are unique and permanent. Classical fingerprint recognition methods that use machine learning often face challenges when processing low-quality samples, which requires forensic experts' assistance. The use of deep learning, which overcomes some of the limitations of classical methods, is becoming increasingly popular, but there are still too few developed solutions in this field.
In this thesis, we developed a model based on Siamese neural networks (SNN) combined with the ResNet34 architecture, enabling us to efficiently compare fingerprints in latent space. We further enhanced the basic model by integrating spatial transformer networks (STN), which ensure rotational invariance, and incorporating domain knowledge about minutiae, adding additional relevant information to the process. We evaluated the methods on several publicly available datasets, where our model achieved a higher level of accuracy compared to some classical methods, confirming the potential of deep learning in the field of fingerprint identification.
|
