Finger vein verification using a Siamese CNN
Finger vein verification using a Siamese CNN
复制标题
使用 Siamese CNN 验证手指静脉
DOI：
10.1049/iet-bmt.2018.5245
复制
发表时间：
2019-09-01
期刊：
影响因子：
2
通讯作者：
Deng, Feiqi
中科院分区：
文献类型：
--
作者：
Tang, Su;Zhou, Shan;Deng, Feiqi
手指静脉识别以其独特的优势受到了越来越多的关注。然而，大多数现有的算法依赖于手工制作的功能，使得它们对手指旋转和偏移的鲁棒性较差。为了缓解这些问题，作者提出了一种新的方法来提取更多的歧视性特征的手指静脉图像。首先，面对训练数据不足的问题，他们采用了重图像增强策略，并开发了一种基于预训练权重的卷积神经网络（CNN）。其次，针对手指静脉验证的特点，他们构建了一种Siamese结构，并结合改进的对比损失函数来训练上述CNN，有效地提高了网络的性能。最后，考虑到在嵌入式设备上部署上述CNN的可行性，他们构建了一个轻量级的CNN，具有深度可分离卷积，并采用知识蒸馏方法从基于预训练权重的CNN中学习知识，这使得它很小但很有效。实验结果表明，轻量级CNN的大小缩小到基于预训练权重的CNN的1/6，而其在MMCBNU_6000，FV-USM和SDUMLA-HMT数据集上的相等错误率分别为0.08，0.11和0.75%，与基于预训练权重的CNN几乎保持相同，并超过了最先进的方法。
Finger vein verification has received more attention recently due to its unique advantages. However, most existing algorithms rely on handcrafted features, making them less robust to finger rotation and offsets. To alleviate these problems, the authors propose a novel method to extract more discriminative features from finger vein images. First, facing the issue of insufficient training data, they adopt a heavy image augmentation strategy and develop a pretrained-weights based convolutional neural network (CNN). Second, focusing on the characteristics of finger vein verification, they construct a Siamese structure combining with a modified contrastive loss function for training the above CNN, which effectively improves the network's performance. Finally, considering the feasibility of deploying the above CNN on embedded devices, they construct a lightweight CNN with depthwise separable convolution and adopt a knowledge distillation method to learn the knowledge from the pretrained-weights based CNN, which makes it small but effective. The experimental results show that the size of the lightweight CNN shrinks to 1/6th of the pretrained-weights based CNN, while its equal error rates achieved in the MMCBNU_6000, FV-USM and SDUMLA-HMT datasets are 0.08, 0.11 and 0.75% respectively, which nearly stays the same with the pretrained-weights based CNN and surpasses state-of-the-art methods.