Segmenting Fingerprints using Deep Learning: A Step towards Touchless Fingerprinting

Fingerprints are considered as one of the most important biometric features of the human body. However, the emergence of COVID-19 has highlighted the limitations of touch-based fingerprint recognition systems, leading to a growing need for touchless alternatives. Additionally, latent fingerprints, presence of dirt on sensor plate only adds to the list of already concerning problems of touch-based fingerprint recognition system. A solution to the same is the introduction of touchless fingerprint technology. Segmentation is a critical step in the touchless fingerprint recognition process, and it has a direct impact on the system's accuracy and reliability. In recent years, deep learning has made significant advancements in various fields like autonomous driving, NLP and many more. In this literature, we explore the importance of segmentation in touchless fingerprint recognition systems and present a workflow based on deep learning. Proposed work semantic segmentation using U-Net with Batch Normalization capable of segmenting distal phalanx for a touchless fingerprint recognition system. The aim is to assess the performance of the model using different loss functions and find the best performing model variation with BCE loss function is further used to test the potency of the model and is tested on different parameters. The model is compared with different segmentation models to empirically understand, where the model stands against state-of-the-art segmentation models. It is observed that the model performs best when BCE loss function is used. This model variation is further tested using different dataset sizes and threshold values with the best mIoU values being obtained at a threshold of 0.8.