摘要
由于现有虹膜数据集样本量较少,利用传统卷积神经网络训练易产生过拟合问题,为此利用嵌入了迁移学习的卷积神经网络对虹膜数据集进行训练,以提升识别准确率。首先对虹膜图像进行预处理,具体用Viterbi算法和橡胶板模型分别对虹膜数据集的眼部区域中的虹膜区域进行分割和归一化,得到规整的虹膜图像集;然后将添加迁移学习和传统的3种卷积神经网络(GoogleNet、VGG、ResNet)在归一化后的虹膜数据集中训练,得到虹膜特征向量;最后用softmax分类器进行N分类。试验结果表明,添加迁移学习的网络模型解决了虹膜数据集训练的过拟合问题,相比传统方法,在4大虹膜数据集(CASIA-Iris-Interval、CASIA-Iris-Lamp、CASIA-Iris-Thousand、IITD Database)上的识别准确率均有大幅提升,最高可达9.3%,其中,ResNet网络模型在各数据集上的表现最优。通过与现有虹膜识别算法的对比,不仅准确率要高于后者,而且平均识别时间和平均训练时间都有大幅度的缩减,证明了该方法的高效性和鲁棒性。
Because the existing iris dataset has a small sample size,the traditional convolutional neural network training is easy to produce over-fitting problem.For this reason,the iris dataset is trained by the convolutional neural network embedded with Transfer Learning to improve the recognition accuracy.The method firstly preprocesses the iris image,and specifically uses the Viterbi algorithm and the Rubber Sheet Model to respectively focus on the eyes of the four iris data sets.The iris region in the region is segmented and normalized to obtain a regular iris image set.Then,the three convolutional neural networks(GoogleNet,VGG,ResNet)with Transfer Learning and tradition are added to train in the normalized iris data set to obtain the iris feature vector.Finally use the softmax classifier for N classification.The experimental results show that the network model with Transfer Learning solves the over-fitting problem of iris dataset training.Compared with the traditional method,the recognition accuracy on the four iris data sets(CASIA-Iris-Interval,CASIA-Iris-Lamp,CASIA-Iris-Thousand I,ITD Database)has been greatly improved,up to 9.3%.Among them,the ResNet network model performs best on each data set.Compared with the existing iris recognition algorithm,not only the accuracy rate is higher than the latter,but also the average recognition time and the average training time are greatly reduced,which proves the efficiency and robustness of the method.
作者
赵勇
雷欢
马敬奇
肖任翔
张寿明
Zhao Yong;Lei Huan;Ma Jingqi;Xiao Renxiang;Zhang Shouming(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650093,China;Guangdong Institute of Intelligent Manufacturing,Guangdong Key Laboratory of Modern Control Technology Guangdong Open Laboratory of Modern Control&Optical,Guangzhou 510070,China;College of Electrical and Information Engineering,Hunan University,Changsha,410082,China)
出处
《电子测量技术》
2020年第9期114-120,共7页
Electronic Measurement Technology
基金
国家自然科学基金(61364022)
广东省科学院青年科技工作者引导专项(2019GDASYL-0105066)资助