High-Speed Dorsal Hand Vein Recognition Using Siamese Lightweight Neural Network

Dorsal hand vein recognition possesses unique advantages such as liveness detection and high stability in identity recognition. However, high computational demands and the need to retrain the model as the number of registered users increases, make it challenging to fully apply current deep learning research results to most edge devices. To address this limitation, MSGNet (Modified Siamese GhostNet) in this paper utilizes siamese GhostNet backbones to reduce computational demands and enhance adaptability to dynamic datasets. Additionally, optimizations such as the introduction of a multi-scale convolution module improve accuracy and robustness. A dataset comprising 118 subjects was constructed for this paper. Experimental results showed that the model achieved an average matching time of 25.61 ms, a recognition rate of 98.82%, and an equal error rate of 1.60%. These results outperforms existing lightweight dorsal hand vein recognition algorithms, establishing MSGNet as a state-of-the-art solution for practical deployment.