Fast High-Resolution Fingerprint Recognition using Domain-Knowledge Infused Global Descriptors

High-resolution fingerprint recognition is mainly centred around local descriptors created using pore patches. Although these methods provide good verification performance, they are not well-suited for identification due to poor computational performance and variable and large template size caused by the variable number of useful pore patches. We present a deep learning model that overcomes this problem by learning to generate a fixed-sized global descriptor while also taking into account the finer level-3 features by infusing domain knowledge using a multi-task architecture. Our approach employs a CNN with two branches simultaneously trained to generate descriptors and pore-intensity maps. We have augmented a publicly available dataset (IITI-HRF) for performance evaluation. Our method compares favorably to the state-of-the-art in terms of accuracy, while being significantly faster ($\sim 24\times$ for verification and $\sim 518000\times$ for identification) and having a smaller template size.