Download files
Abstract
It is extremely challenging to identify a partial fingerprint against a large database due to the inability to narrow down the candidate list for partial fingerprint verification. Furthermore, the traditional capture of fingerprints based on the contact of the finger on a solid plane results in partial or degraded images. In this thesis, we aim to devise effective indexing schemes for partial fingerprint identification against very large scale databases. Furthermore, we have also acquired databases and developed identification techniques for the 3D images of fingerprints that have been generated by the new generation of touchless live scan devices.
For partial fingerprint indexing, we propose to combine both local feature and global feature. We design some novel features of minutiae triplets in addition to some commonly used features to constitute the local minutiae triplet features. We then propose to combine a reconstructed global feature and local minutiae triplet features to improve the performance of partial fingerprint indexing. Specifically, the minutiae triplet based indexing scheme and the FOMFE coefficients based indexing scheme are applied separately to generate two candidate lists, then a fuzzy-based fusion scheme is designed to generate the final candidate list for matching.
We have collected a multi-sensor fingerprint database to investigate the 3D fingerprint biometric comprehensively. It consists of 3D fingerprints as well as their corresponding 2D fingerprints captured by two commercial fingerprint scanners from 150 subjects in Australia. Also, we have tested the performance of 2D fingerprint verification, 3D fingerprint verification, and 2D to 3D fingerprint verification. In addition, the database has been released publicly for research purposes since 2015.
For multi-sensor fingerprint indexing, we propose a finer hash bit selection method based on Locality-Sensitive Hashing (LSH) and Minutia Cylinder-code (MCC). That is, we divide the hash bit vectors, selected by LSH using a sliding window, into finer sub-vectors with certain fixed length, and then convert these sub-vectors into numerical approximations for MCC indexing. Also, we take into consideration another feature - the single maximum collision for indexing and fuse the candidate lists produced by both indexing methods to produce the final candidate list.