Efficient hand vein recognition using local keypoint descriptors and directional gradients

This paper proposes a computationally efficient palm and wrist vein biometric system through finely tuning computer-vision algorithms. In particular, a comprehensive analysis of the scale-invariant feature transform (SIFT) and speeded-up robust features (SURF) keypoint descriptors was conducted along with a novel idea of a score-based fusion of directional image derivatives to achieve outstanding recognition results. The work demonstrates that appropriate vein image processing, keypoint extraction, optimal matching metrics, and combination of classification scores from a group of directional gradients lead to robust and stable vein recognition. It was shown through experimental analysis that the developed biometric system outperforms all state-of-the-art results other than deep learning methods on the two public hand vein databases (VERA and PUT). Moreover, an absolute 100% recognition for the PUT palm dataset was achieved without using deep learning. The proposed method is more suitable for embedded implementation compared to deep learning algorithms, with only a slight penalty in performance compared to deep learning architectures.