Vision transformers for fingerprint embedding generation: Evaluation on CASIA dataset

Fingerprint recognition systems traditionally rely on minutiae-based or CNN-based feature extraction methods. This paper investigates the application of Vision Transformers (ViTs) for generating discriminative fingerprint embeddings. We employ standard ViT architecture with 16×16 patch size, originally designed for natural image classification, and adapt it for fingerprint biometric encoding using triplet loss optimization. Our approach treats fingerprint images as sequences of patches, leveraging the self-attention mechanism of transformers to capture both local ridge patterns and global fingerprint structure. We evaluate our method on the CASIA Fingerprint Database, conducting verification experiments with 114 genuine pairs and 114 impostor pairs. The results demonstrate strong performance, achieving a ROC AUC of 0.9899 and an Equal Error Rate (EER) of 4.82%. These results indicate that Vision Transformers, without fingerprint-specific architectural modifications, can effectively learn discriminative embeddings for fingerprint recognition. The success of this approach suggests that transformer-based architectures represent a viable alternative to conventional methods, opening new directions for biometric feature extraction using attention mechanisms.