Touchless Fingerprint Classification Using Deep Learning

Yong, Jeremy Yi Jie (2026) Touchless Fingerprint Classification Using Deep Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text JEREMY_YONG_YI_JIE_Full Text.pdf Restricted to Registered users only Download (10MB)

Abstract

This study presents a Vision Transformer based framework for touchless fingerprint classification as a hygienic and non-intrusive alternative to contact-based biometric recognition systems. A structured preprocessing pipeline was developed incorporating contrast enhancement through Contrast Limited Adaptive Histogram Equalization (CLAHE), Gaussian noise reduction, and region-of-interest (ROI) extraction to standardise input quality under variable capture conditions. Seven deep learning models spanning both Convolutional Neural Network (CNN) and Vision Transformer architectures were evaluated under identical preprocessing and training conditions, comprising EfficientNetV2, MobileNetV4, ResNet50, SwinV2, ViT Base, DeiT Tiny, and MobileViT. While CNN based models achieved the highest baseline classification accuracy, SwinV2 demonstrated competitive performance among transformer based architectures and was selected for further refinement due to its hierarchical shifted-window attention mechanism and suitability for small domain-specific datasets. Refinement through a modified two-layer MLP classification head and a staged unfreezing training strategy yielded a final classification accuracy of 93.51% across five fingerprint pattern classes. The introduction of mixup augmentation was identified as a critical factor in improving model generalisation, increasing accuracy from approximately 80% to above 90% across evaluated architectures. Analysis of misclassified samples revealed that most errors arise from intrinsic structural similarity between pattern classes, particularly between arch and loop formations, rather than model instability. The outcomes of this study advance touchless biometric recognition and align with the United Nations Sustainable Development Goals on good health, innovation, and strong institutions