Learning a Fixed-Length Fingerprint Representation

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Traditional fingerprint matching relies on variable-length minutiae representations.
• Minutiae-based methods face challenges with computational cost, security, and accuracy in low-quality prints.

Purpose of the Study:

• To introduce DeepPrint, a deep learning network for generating compact, fixed-length fingerprint representations.
• To demonstrate the advantages of DeepPrint over existing minutiae-based approaches.

Main Methods:

• Developed a deep network architecture incorporating fingerprint domain knowledge (alignment, minutiae detection).
• Extracted fixed-length fingerprint representations of 200 bytes.
• Benchmarked DeepPrint against commercial systems (Verifinger, Innovatrics) using NIST SD4 dataset.

Main Results:

• DeepPrint achieved comparable rank-1 accuracy (98.80%) to top commercial matchers on a large dataset.
• DeepPrint significantly reduced search time (0.3 seconds vs. 27 seconds).
• The compact representation offers improved security and discriminative power, especially for low-quality fingerprints.

Conclusions:

• DeepPrint provides the most compact and discriminative fixed-length fingerprint representation to date.
• The method offers a faster, more secure, and robust alternative for fingerprint identification.