Learning compact binary quantization of Minutia Cylinder Code

With explosive growth in fingerprint database, Automatic Fingerprint Identification System (AFIS) has become more difficult than ever. Consequently, it is necessary to get an effective and discriminative fingerprint feature binary representation. In this paper, we firstly analyze the characteristic of Minutia Cylinder Code (MCC) representation to find that it is strongly bit-correlated and with a lossy binary quantization. Accordingly, we propose an optimization model to learn a feature projection matrix resulting in dimensionality reduction as well as diminishing quantization loss. Eventually, the real-valued version of MCC is learnt to get Compact Binary Minutia Cylinder Code (CBMCC) with balanced independent property and minimal binary quantization loss. The performance test shows that CBMCC is effective and discriminative as it has maximum intra-bit variance while minimum inter-bit correlation. Furthermore, numerous experiments on public databases demonstrate that CBMCC is advantageous for fingerprint retrieval since it achieves a high correct index performance with a fairly low penetration rate.