Vision Transformers for Vein Biometric Recognition

In October 2020, Google researchers present a promising Deep Learning architecture paradigm for Computer Vision that outperforms the already standard Convolutional Neural Networks (CNNs) on multiple image recognition state-of-the-art datasets: Transformers (ViTs). Based self-attention concept inherited from Natural Language Processing (NLP), this new structure surpasses CNN classification task ImageNet, CIFAR-100, and VTAB, among others, when it is fine-tuned (Transfer Leaning) after previous pre-training larger datasets. work, we confirm theory move one step further over structures applied Vascular Biometric Recognition (VBR): to best of our knowledge, introduce first time pure pre-trained in evolving biometric modality address challenge limited number samples VBR For purpose, ViTs have been trained extract unique features ImageNet-1k ImageNet-21k then four main existing variants, i.e., finger, palm, hand dorsal, wrist vein areas. Fourteen vascular datasets used perform identification previously mentioned modalities, based True-Positive Identification Rate (TPIR) 75-25% train-test sets obtaining following results: HKPU (99.52%), FV-USM (99.1%); Vera (99.39%), CASIA (96.00%); Bosphorus (99.86%); PUT-wrist (99.67%), UC3M-CV1+CV2 (99.67%). Furthermore, UC3M-CV3: hygienic contactless database collected smartphones consisting 5400 images 100 different subjects. The results show Transformer’s versatility under Transfer reinforce Network paradigm.