Finger-vein feature extraction method based on vision transformer

Lu, Zhiying; Wu, Runhao; Zhang, Jianfeng

doi:10.1117/1.jei.31.4.043010

Cited by 2 publications

(1 citation statement)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Zhao et al [22] proposed a lightweight convolutional neural network based on a central loss function and dynamic regularization, which not only reduced the false recognition rate but also accelerated the convergence speed. Lu et al [23] proposed a lightweight model based on Vision Transformer (ViT) T2T-ViT, where tokens input into ViT are subjected to Reshape and Soft Split operations to reduce the dimensionality of tokens, making the overall model more lightweight, but the model loses a certain amount of accuracy. Although all of the above lightweight networks have achieved good recognition results, the reduction in the number of network layers also represents the acquisition of less global information, which can only be based on the shallow network extracted edges, corners, points, and some vein texture change information to discriminate, the global information is not enough to grasp, and it is easy to ignore some important details of the information is not conducive to improving the recognition performance.…”

Section: Introductionmentioning

confidence: 99%

FV-DMHN: Dual Multi-Head Network for Finger Vein Recognition

An,

Ren,

Tao

2024

IEEE Access

View full text Add to dashboard Cite

Deep learning-based finger vein image recognition methods usually suffer from high complexity, insufficient global information extraction, and overfitting. The use of lightweight networks can significantly reduce the accuracy owing to the reduction in model parameters. For this reason, this paper proposes a Dual Multi-Head neural Network for Finger Vein Recognition (FV-DMHN), which combines the Multi-Head Self-Attention (MHSA) mechanism with the Multi-Head Convolutional Network(MHCN)to increase the training efficiency of the network while expanding the CNN horizon. The inverted residual structure is also used in the network to enhance the expressive power of the network. The algorithm achieves recognition accuracies of 99.81%, 99.67%, 99.69%, and 99.83% on three publicly available datasets, FV-USM, SDUMLA-HMT, THU-FVFDT2, and self-built datasets FV-SIPL, respectively, with an average equal-error rate of 0.339%, and the recognition time of a single image is only 2.63 ms. The experimental results show that the algorithm is superior to other methods in terms of accuracy and average equal error rate, at the same time, it not only reduces the number of network parameters and computational complexity but also achieves excellent recognition speed.

show abstract

Section: Introductionmentioning

confidence: 99%