Veintr: robust end-to-end full-hand vein identification with transformer

Hand vein identification stands out to be an increasingly popular approach for biometric identification due to its distinctiveness and convenience. While state-of-the-art techniques are able to achieve good performance, they share two common drawbacks: (1) complex preprocessing procedures, e.g., vein enhancement and Region of Interest (ROI) extraction, and (2) vein information loss due to hand ROI partition. To address these issues, we propose VeinTr, an end-to-end full-hand vein identification approach. In particular, our VeinTr consists of three components: a local feature extractor, a lightweight transformer, and a global feature decoder. We first obtain local features via convolution-based ResNet-like blocks. Then the attention mechanism is employed to aggregate global features from local features, which can be then decoded as global hand vein features. Finally, a global feature decoder is applied to generate robust hand features. By doing so, VeinTr is capable of directly extracting robust hand vein features from raw hand vein images. We evaluate our method on CASIA, TPV, and PLUSVein hand vein datasets. Experimental results show that our approach outperforms the state-of-the-art methods and has strong inter-dataset generalization abilities.