Julius Cewers, Kalle Svensson

Abstract: 

Biometric fingerprint recognition is a cornerstone of modern authentication systems, yet conventional convolutional neural network approaches often struggle with challenging cases, subjected to noise and partial overlap. This thesis explores the potential of transformer-based architectures to address these challenges.We design, implement, and evaluate five models on a large-scale dataset of over five million annotated fingerprint pairs: ViT-Small, a lightweight Vision Transformer trained from scratch; ViTLarge, a high-capacity variant boosting and weighting implemented in training; ViT-Large-Without, with the same structure as ViT-Large but without boosting or weighting; ViT-Large-Combined that combines the approaches of the earlier two; and Swin Transformer, which employs hierarchical, shifted-window attention for efficient multi-scale feature extraction. To simulate real-world variability, we include noise reflecting wet, dry, and environmental conditions into our training. Our experiments demonstrate that especially ViT-Large-Without performs well, achieving high performance in the relatively short development time. Beyond accuracy gains, we analyze inference latency and GPU utilization, revealing that transformer models can be deployed with practical resource requirements. These findings demonstrate that a high-capacity Vision Transformer trained without complex weighting schemes strikes an optimal balance between robustness, accuracy, and computational efficiency. ViT-Large-Without emerges as the most promising candidate for real-world biometric systems, challenging conventional assumptions that elaborate training strategies are necessary for state-of-the-art fingerprint matching.