Finger Vein Recognition Based on ResNet With Self-Attention

Abstract

To solve the problem of low accuracy and high computational resource consumption in finger vein recognition, a finger vein recognition model based on ResNet with self-attention (FV-RSA) is proposed. This model combines global focusing ability of self-attention mechanism and local feature extraction ability of CNN, which improves recognition accuracy. To reduce the number of parameters and floating-point operations, self-attention and convolution share linear projections by pointwise convolution. Self-attention and CNN are fused in the convolution and self-attention (CASA) block connected by skip connection to avoid gradient vanishing or gradient exploding. During the training phase, we use a variable learning rate with cosine annealing to avoid falling into local optimum. Experiments show that the method works well on the public database, which can not only improve the accuracy, but also reduce the number of parameters and computational complexity.

Keywords

• Finger vein recognition
• deep learning
• ResNet
• self-attention
• variable learning rate