Fingerprint indexing with pose constraint

Fingerprint indexing is critical for identifying fingerprints efficiently in large-scaled fingerprint databases. The state-of-the-art indexing accuracies are achieved by minutiae-based indexing approaches. A major difference of these approaches is how they deal with fingerprint registration problem. Some of the approaches do not use registration, while others perform relative/pairwise registration (e.g., based on mated minutiae between two fingerprints). However, the former approach is not accurate, since without geometric constraints, many false matches can be found even for non-mated fingerprints. The latter is not efficient for large databases since the relative registration step has to be performed for each pair of fingerprints. It is desirable to develop an absolute registration approach, which can register any single fingerprint into a common coordinate system, and therefore can address efficiency and accuracy problems simultaneously. In this paper, we proposed a fingerprint pose estimation algorithm which can register fingerprints into a common finger coordinate system. Fingerprint pose estimation problem is viewed as a two-class classification problem and approached by sliding window classifiers trained on labeled data. Ridge orientation information is used as features instead of singularities which are often affected by noise. With the pose estimation, we are able to refine the matched minutiae with global spatial constraint. By combining the proposed pose estimation algorithm with an improved Locality Sensitive Hashing algorithm for MCC descriptor, our indexing system outperformed previous state-of-the-art on widely used databases and its scalability was tested on a large database with one million fingerprints.