Large-scale fingerprint identification on GPU

Abstract

This paper proposes a new parallel algorithm to speed up fingerprint identification using GPUs. A careful design of the algorithm and data structures, guided by well-defined optimization goals, yields a speed-up of 1946× over a baseline sequential CPU implementation and of 207× over a CPU implementation optimized with SIMD instructions. The proposed algorithm enables a medium-scale AFIS (Automated Fingerprint Identification System) to run on a simple PC with four Tesla C2075 GPUs. On a benchmark with 250000 fingerprints and 100000 queries, the proposed system yields state-of-the-art biometric accuracy with a throughput of more than 35million fingerprint matches per second. The proposed approach can be easily scaled-up, thus making possible the implementation of a large-scale AFIS (i.e., with a database of hundred million fingerprints) on inexpensive hardware.