基于静脉关键特征和AdaFace损失的轻量级指静脉识别算法

基于深度学习的指静脉识别方法通常需要大量的计算资源，限制了其在嵌入设备上的推广和普及，采用轻量级网络又面临模型参数减少导致准确率下降的问题，为此提出一种基于指静脉关键特征和AdaFace损失的轻量级识别算法。在MicroNet框架中，首先提出一种FMixconv卷积来替代原网络中的深度卷积，减少参数的同时可以获得静脉特征的多尺度信息；其次引入轻量级注意力模块CA模块，从空间和通道上聚焦于静脉特征的关键信息；最后在损失函数中加入AdaFace损失，通过特征范数对图像质量进行评价，以减少图像质量下降对训练的影响。该算法在SDUMLA-HMT、FV-USM和自建数据集上的识别准确率达到99.84%、99.39%和99.42%,而参数量仅有0.82 M。实验结果表明，该算法在准确率和参数量大小上均领先于其他方法。

Research on lightweight recognition algorithm based on key features of finger vein and AdaFace loss

Finger vein recognition methods based on deep learning usually require a large amount of computing resources, it limits their promotion and popularization on embedded devices. The adoption of lightweight network faces the problem of decreasing accuracy due to the reduction of model parameters. Therefore, this paper proposed a lightweight recognition algorithm based on key features of finger vein and AdaFace loss. In the MicroNet network framework, firstly, this paper proposed FMixconv convolution to replace the deep convolution in the original network, which could obtain multi-scale information of vein features while reducing parameters. Secondly, the method used a lightweight attention module, CA module, to focus on key information of venous characteristics from space and channel. Finally, the algorithm added AdaFace loss into the loss function, through the characteristics of the norm to evaluate image quality, to reduce the impact of image quality degradation on training. The recognition accuracy of the proposed algorithm on SDUMLA-HMT, FV-USM and self-built datasets reached 99.84%, 99.39% and 99.42%, while the number of parameters was only 0.82M. Experimental results show that the proposed network is ahead of other methods in accuracy and parameter size.