摘要
针对手指静脉与手指关节纹的数据样本小且识别准确率易受各自固有属性限制以及非注册用户对系统识别准确率影响较大等问题,提出一种基于迁移学习的带拒绝识别阈值的手指静脉与手指关节纹共同决策同一主体的双模态分数级融合识别方法。对二者数据集进行数据扩充和图像尺寸调整;使用经ImageNet海量数据集训练后的Vgg19、Inceptionv3、Xception以及Resnet分别在二者数据集上进行参数调优;应用调优后的新模型进行分类识别,得到各自的匹配分数,再进行分数级融合,融合后的匹配分数与拒绝识别阈值比较,再进行最终的决策。该方法在公开数据集中识别准确率均可达99%,较各自单模态在各个网络中的识别准确率提高0.33%~15%不等。实验结果表明,采用迁移学习方法对指静脉与指关节纹进行分数级融合能够有效提高系统的识别准确率。
Aiming at the problems of small data samples of finger vein and finger knuckle print,and the identification accuracy was easily limited by their inherent attributes,as well as the influence of non-registered users on the identification accuracy of the system,this paper proposes bimodal fractional fusion identification method based on transfer learning with rejection identification threshold for finger vein and finger knuckle print.Data expansion and image size adjustment were performed for both datasets.We used Vgg19,Inception v3,Xception and Resnet trained by ImageNet massive datasets to optimize parameters on the two datasets respectively.Finally,the optimized new model was applied to classify and identify,and the matching scores were obtained and the scores were fused.After the fusion,the matching scores were compared with the rejection recognition thresholds,and the final decision was made.The identification accuracy of the proposed method can reach 99%in the open dataset,which is 0.33%~15%higher than the identification accuracy of each single mode in each network.The experimental results show that the score level fusion method of finger vein and finger knuckle print using transfer learning can effectively improve the identification accuracy of the system.
作者
陶志勇
冯媛
林森
Tao Zhiyong;Feng Yuan;Lin Sen(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,Liaoning,China;Fuxin LiXing Technology Co.,Ltd,Fuxin 123000,Liaoning,China)
出处
《计算机应用与软件》
北大核心
2019年第12期162-168,183,共8页
Computer Applications and Software
基金
辽宁省博士启动基金项目(20170520098)
辽宁省普通高等教育本科教育改革研究项目(551610001095)
辽宁省教育厅一般项目(LJ2017QL013)
关键词
手指静脉识别
手指关节纹识别
迁移学习
分数级融合
Finger vein identification
Finger knuckle print identification
Transfer learning
Score level fusion
