Regression Via Classification for Fingerprint Orientation Estimation

Abstract

Estimating the direction in which the ridges and valleys of the fingerprint pattern are aligned often serves as a pivotal first step in fingerprint recognition systems. The ridge orientation map is a fundamental reference for subsequent processing stages, such as image enhancement, feature extraction, and matching. Therefore, its accuracy is essential to achieve high recognition rates. Ridge orientation estimation entails a regression problem since the task is to estimate an angle between 0 degrees and 180 degrees for each sub-region in the fingerprint image. However, the majority of the approaches in the literature pivot towards framing this regression task as a classification problem. This paper systematically analyzes the regression via classification methodology for fingerprint orientation estimation, exploring various discretization and encoding strategies. Specifically, we examine single and multiple discretization schemes designed to ensure that resulting bins maintain uniform length or uniform probability or are allocated randomly, paired with one-hot, ordinal, and cyclic encoding techniques. Our experiments are conducted on the FOE-TEST database from FVC-onGoing, the sole publicly available fingerprint orientation dataset. The findings highlight the efficacy of cyclic encoding over the one-hot encoding prevalent in prior research, while equal-length and equal-probability discretization strategies yield comparable results.