Exploring machine learning techniques for fingerprint ridge counting

Abstract

Ridge counting is an important feature in standardized fingerprint templates, supporting interoperability, improving matching accuracy, and providing a transparent and interpretable measure to validate automated decisions. However, traditional ridge counting techniques that rely on binarization and skeletonization are prone to artifacts, limiting their accuracy and robustness. In this study, we propose and analyze machine learning methods for ridge counting based on different architectures paradigms such as fully connect residual networks, convolutional and Transformer techniques, which operate directly on raw grayscale fingerprint images. As part of this study, we introduce a benchmark dataset comprising 23,724 of manually annotated ridge counts from 50 subjects. Our experiments show that convolutional and Transformer-based models achieved the highest accuracy, with exact ridge count match rates of 96.6% and 95.6%, respectively, outperforming classical techniques and a commercial solution top ranked in the NIST MINEX III evaluation. Additionally, we investigate the influence of ridge counting on fingerprint matching performance using the FVC2002 benchmark, with additional cross-dataset tests on Fingerprint Verification Competition 2004 DB1 A (FVC2004) and SD301. The source codes, ridge count dataset, and trained models are available on GitHub.