Automated vein verification using self-attention-based convolutional neural networks

Vein-based biometric traits have been regarded as trustworthy for biometric applications. With technical advances in deep learning, verification performance has started to be improved in these applications to increase trust level in daily life by providing usage convenience and user satisfaction. In this study, the effect of self-attention mechanism on convolutional neural networks for the performance of finger-vein and hand dorsal vein verification was investigated using an open-set protocol. To provide generalizability to the trained model, self-attention-based convolutional neural networks were used rather than existing architectures and pre-trained models. With the architecture that uses residual blocks and self-attention mechanism, a fair verification performance was suggested. Verification performance was assessed on DHVI-DB and Bosphorus hand dorsal vein datasets and SDUMLA and PolyU-F finger-vein datasets in terms of equal error rate using the distance between feature vectors through the existing and the proposed distance measures. The obtained equal error rates for hand dorsal vein datasets DHVI-DB1, DHVI-DB2, and Bosphorus are 2.17, 2.21, and 18.33, respectively and for finger-vein datasets SDUMLA and PolyU-F, are 1.65 and 10.64, respectively. Moreover, 4 different loss functions were used throughout the conducted experiments to see the discriminative ability of the proposed network for vein verification. The experimental results on these datasets indicate the potential effectiveness of the self-attention mechanism on automated vein verification.

Automated summary

This study investigates the effect of the self-attention mechanism on the performance of finger-vein and hand dorsal vein verification using convolutional neural networks. The experimental results suggest the potential effectiveness of the self-attention mechanism for automated vein verification.