Fingerprint Synthesis: Evaluating Fingerprint Search at Scale

Abstract

A database of a large number of fingerprint images is highly desired for designing and evaluating large scale fingerprint search algorithms. Compared to collecting a large number of real fingerprints, which is very costly in terms of time, effort and expense, and also involves stringent privacy issues, synthetic fingerprints can be generated at low cost and does not have any privacy issues to deal with. However, it is essential to show that the characteristics and appearance of real and synthetic fingerprint images are sufficiently similar. We propose a Generative Adversarial Network (GAN) to generate 512X512 rolled fingerprint images. Our generative model for rolled fingerprints is highly efficient (12ms/image) with characteristics of synthetic rolled prints close to real rolled images. Experimental results show that our model captures the properties of real rolled fingerprints in terms of (i) fingerprint image quality, (ii) distinctiveness and (iii) minutiae configuration. Our synthetic fingerprint images are more realistic than other approaches.