Dense Fingerprint Registration via Displacement Regression Network

Dense registration of fingerprints provides pixel-wise correspondences between two fingerprints, which is beneficial for fingerprint mosaicking and matching. However, this problem is very challenging due to large distortion, low fingerprint quality and lack of distinctive features. The performance of existing dense registration approaches, such as image correlation and phase demodulation, are limited by manually designed features and similarity measures. To overcome the limitations of these approaches, we propose a dense fingerprint registration algorithm through convolutional neural network. The key component is a displacement regression network (DRN) that can regress pixel-wise displacement field directly from coarsely aligned fingerprint images. Training ground-truth data is automatically generated by an existing dense registration algorithm without tedious manual labelling. We also propose a multi-scale matching score fusion method to show the application of the proposed registration algorithm in improving fingerprint matching accuracy. Experimental results on FVC2004 DB1_A and DB2_A, and Tsinghua Distorted Fingerprint (TDF) database show that our method reaches state-of-the-art registration performances.