Deep Fingerprint Classification in a Low-cost Environment

Fingerprint classification is plays an essential role in human identification mechanisms. For systems with a big fingerprint database, an accurate fingerprint classification algorithm can significantly reduce fingerprint searching and matching time. In this respect, developing an accurate algorithm based on deep learning may be an ideal solution. However, issues on the computational complexity and inadequacy to invest on expensive graphic processing units (GPUs) make the development task becomes challenging. In this paper, we present a deep fingerprint classification algorithm that can achieve better performance compared to the previously reported in the literature without a requirement to invest in GPUs. For these purposes, we modify and train the pre-trained model DenseNet-121 on publicly available GPUs offered by Google's Colaboratory. The performance of our framework is evaluated on images from the NIST-4 database. In the five-class classification problem, the proposed algorithm is superior to all the competing algorithms, with an average accuracy of 96.3%. When used for the four-class classification scheme, our algorithm can maintain an average accuracy of 97.7%, significantly better than the competing algorithms. The obtained results and the fact that no compulsion to invest in expensive GPUs suggest that our algorithm would be useful for human recognition systems.