信息技术的进步催生了生物特征识别逐渐替代传统身份验证方法，尤其关注卫生、安全的掌静脉识别，然而在计算资源受限的情况下保持识别性能仍然是一项挑战。近年来，虽然深度学习架构Vision Transformer在模型性能上取得显著进展并在掌静脉识别领域逐渐得到应用，但是也因参数量问题限制了其适用范围。该文提出了一种手掌静脉识别算法(NAM-EfficientViT Based Palm Vein Recognition，NEPVR)，采用了 EfficientViT作为深度学习的高效轻量化网络以减少参数量的规模，并结合归一化注意力机制加强图像在通道和空间维度上对重要细节特征的提取，进而保持良好的识别性能。此外，NEPVR还融合了交叉熵和三元组损失函数作为在网络训练中的综合损失函数，以提高识别性能和模型收敛的稳定性。实验结果表明:将掌静脉信息编码为512维特征向量的方法识别性能最佳;在PolyU、CASIA与TongjiU数据集上进行的评估中，等误差率(EER)分别达到了 0。067％、0。150％与0。085％，充分证明了该算法的有效性。
A Lightweight Palm Vein Recognition Algorithm NEPVR
Advances in information technology have led to the gradual replacement of traditional authentication methods by biometrics,with a particular focus on hygienic and secure palm vein recognition,yet maintaining the recognition performance in the presence of limited computational resources remains a challenge.In recent years,although Vision Transformer,a deep learning architecture,has made significant progress in model performance and has been gradually applied in palm vein recognition,which has also been limited in its ap-plicability due to the number of parameters.We propose a palm vein recognition algorithm(NAM-EfficientViT-based Palm Vein Rec-ognition,NEPVR),which adopts EfficientViT as an efficient lightweight network for deep learning to reduce the scale of the number of parameters,and combines with normalization-based attention module to enhance the extraction of important detailed features in the image in both the channel and spatial dimensions,and thus maintains excellent recognition performance.In addition,NEPVR incorporates cross entropy and triple loss function as the combined loss function in network training to improve the recognition performance and the stability of model convergence.The experimental results show that the method of encoding palm vein information into 512-dimensional feature vectors has the best recognition performance.In the evaluations conducted on PolyU,CASIA and TongjiU databases,the Equal Error Rate(EER)reaches 0.067％,0.150％and 0.085％respectively,which fully proves the effectiveness of the proposed algorithm.
万方数据
维普
