Learning Binary Descriptors for Fingerprint Indexing

Fingerprint indexing is studied widely with the real-valued features, but few works focus on the binary feature descriptors, which are more appropriate to retrieve fingerprints efficiently in the largescale fingerprint database. In this paper, the binary fingerprint descriptor (BFD), which is an effective and discriminative binary feature representation for fingerprint indexing, is proposed based on minutia cylinder code (MCC). Specifically, we first analyze MCC to find that it has characteristics of the high dimensionality, redundancy, and quantization loss. Accordingly, we propose an optimization model to learn a feature-transformation matrix, resulting in dimensionality reduction and diminishing quantization loss. Meanwhile, we also incorporate the balance, independence, and similarity-preservation properties in this learning process. Eventually, a multi-index hashing-based fingerprint indexing scheme further accelerate the exact search in Hamming space. The experiments on numerous public databases show that the BFD is discriminative and compact and that the proposed approach is outstanding for fingerprint indexing.