摘要
提出一种基于深度残差网络的轻量级指静脉识别算法。首先,以ResNet34为基础,使用深度可分离卷积代替传统卷积,加入SE(Squeeze and Excitation)注意力机制模块来提取手指静脉空间域上的细节特征,并引入宽度缩放因子,进一步压缩网络;其次,在训练中引入教师-学生网络模式,对轻量级深度残差网络进行知识蒸馏训练,并使用知识蒸馏损失、CurricularFace和交叉熵损失对网络进行联合监督,解决了轻量级深度残差网络因学习参数量较少引起的性能下降问题。分别在FV-USM数据集、Lab-Normal数据集和Lab-Special数据集上进行仿真实验,结果表明,同基于轻量级网络MobileFaceNet的识别算法相比,提出的算法有效提高了零误识识别率和Top1排序性能。
A lightweight finger vein recognition algorithm based on deep residual network is proposed.Firstly,based on ResNet34,use depth-separable convolution instead of traditional convolution,add SE(Squeeze and Excitation)attention mechanism module to extract detailed features in the finger vein space domain,and introduce width scaling factor to further compress the network;Secondly,the teacher-student network mode is introduced in the training to carry out knowledge distillation training on the lightweight deep residual network,and the network is jointly supervised by the knowledge distillation loss,“CurricularFace”and“cross-entropy loss”.The performance degradation of the lightweight deep residual network caused by the small number of learning parameters is solved.Simulation experiments are carried out on the FV-USM data set,Lab-Normal data set and Lab-Special data set,respectively.The results show that compared with the recognition algorithm based on the lightweight network MobileFaceNet,the proposed algorithm effectively improves the zero-false recognition rate and Top1 sorting performance.
作者
牟家乐
沈雷
刘浩
郑鹏
MOU Jiale;SHEN Lei;LIU Hao;ZHENG Peng(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China)
出处
《杭州电子科技大学学报（自然科学版）》
2023年第2期35-40,46,共7页
Journal of Hangzhou Dianzi University：Natural Sciences
关键词
指静脉识别
网络压缩
知识蒸馏
宽度缩放因子
深度学习
finger vein recognition
network compression
knowledge distillation
width scaling factor
deep learning
