A Deep Learning Based Approach to Perform Fingerprint Matching

Abstract

As we move towards a technological-driven era, the traditional methods of authenticating an individual are becoming redundant and easier to crack. Biometric authentication has emerged as a very promising authentication method as it uses physiological and behavioral characteristics of the human body, which are unique to every individual. To adopt recent developments in deep learning and apply them to biometric authentication, we propose a novel Vision Transformer (ViT) based Siamese network framework for fingerprint matching in a fingerprint authentication system. Our primary focus is holistic, and an end-to-end pipeline has been constructed and implemented using an ensemble of task-specific algorithms to procure the best possible result from the model. We also endeavor to identify specific problems in the application of ViT to fingerprint matching and used two existing approaches, which are Shifted Patch Tokenization (SPT) and Localized Self Attention (LSA), to tackle those shortcomings effectively. We have considered two variations for the model, namely Intermediate-Merge (I-M) Siamese network and Late-Merge (L-M) Siamese network, and tested the performances on the IIT Kanpur fingerprint database. The obtained results in terms of accuracy, True Positive Rate (TPR), and True Negative Rate (TNR) clearly show the significant performance of the proposed technique for fingerprint matching by means of an approach based on deep learning. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.