Effect of resolution and image quality on combined optical and neural network fingerprint matching

Abstract

Abstract This paper presents results on direct optical matching, using Fourier transforms and neural networks for matching fingerprints for authentication. Direct optical correlations and hybrid optical neural network correlation are used in the matching system. The test samples used in the experiments are the fingerprints taken from NIST database SD-9. These images, in both binary and gray-level forms, are stored in a VanderLugt correlator (A. VanderLugt, Signal detection by complex spatial filtering, IEEE Trans. Inform. Theory IT-10 (1964) 139–145). Tests of typical cross correlations and autocorrelation sensitivity for both binary and 8 bit gray images are presented. When Fourier transform (FT) correlations are used to generate features that are localized to parts of each fingerprint and combined using a neural network classification network and separate class-by-class matching networks, 90.9% matching accuracy is obtained on a test set of 200,000 image pairs. These results are obtained on images using 512 pixel resolution. The effect of image quality and resolution are tested using 256 and 128 pixel images, and yield accuracy of 89.3 and 88.7%. The 128-pixel images show only ridge flow and have no reliably detectable ridge endings or bifurcations and are therefore not suitable for minutia matching. This demonstrates that Fourier transform matching and neural networks can be used to match fingerprints which have too low image quality to be matched using minutia-based methods. Since more than 258,000 images were used to test each hybrid system, this is the largest test to date of FT matching for fingerprints.