HQ-finGAN: High-Quality Synthetic Fingerprint Generation Using GANs

Ataher Sams, Homaira Huda Shomee, S. M. Mahbubur Rahman

Abstract: Artificial fingerprint synthesis can broaden the opportunity for researchers by generating a large number of realistic synthetic fingerprints without having to worry about legal issues or privacy concerns. This paper proposes StyleGAN2 and CycleGAN based dual generative adversarial networks (GAN) system, in which StyleGAN2 generates distinct fingerprint skeletons from preprocessed data and CycleGAN transforms these skeletons into realistic fingerprints. This model can generate high-quality 256 by 256 fingerprints that can be turned into a variety of realistic fingerprint styles. Synthesized fingerprints from this model also retain features of real fingerprints that can be used in the related search system. Experimentation of the model includes visual image quality, quantitative image quality, distinctiveness test, and human perception test. The proposed model can produce more realistic high-quality fingerprints in large quantity as compared to previously reported GAN-based systems.