生物模板保护背景下的迁移学习  

Transfer learning in the context of biological template conservation

作　　者：李浩[1] 孙水发[1] LI Hao;SUN Shuifa(School of Electrical and New Energy,China Three Gorges University,Yichang 443000,China)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443000

出　　处：《现代电子技术》2023年第9期78-82,共5页Modern Electronics Technique

摘　　要：基于深度卷积神经网络(CNN)的方法是生物模板保护中较流行的技术。生物特征信息作为人的唯一属性,具有最高安全保护性。目前,大量研究工作正致力于通过更强大的模型架构和更好的学习技术来进一步改进匹配精度。然而,在探索现有深度人脸识别模型的特征提取能力的研究仍然相对较少。文中分析了经典的三种深度学习网络在不同人脸数据集的特征提取能力,具体来说,对比了VGG16、ResNet50、GoogleNet在同一种模板保护下的性能。仿真结果表明:在图像退化的人脸验证VGG16性能优于其他模型;在图像质量高的情况下,ResNet50最优;GoogleNet在面对复杂学习任务性能更强。匹配精度表明迁移学习优于绝大多数特征提取方法。The methods based on deep convolutional neural networks(CNN)are the more popular techniques in biological template protection.Biometric feature information,as the only attribute of a person,has the highest security protectiveness.Currently,a lot of research work is being devoted to further improving the matching accuracy by more powerful model architectures and better learning techniques.However,there are still relatively few studies in the literature exploring the feature extraction capabilities of existing deep face recognition models.The feature extraction capabilities of three classic deep learning networks in different face datasets are analyzed in this paper.The performances of VGG16,ResNet50 and GoogleNet are compared under the protection of the same template.The simulation results show that the performance of VGG16 is better than other models in the face verification of image degradation,ResNet50 is the best in the case of high image quality,and GoogleNet has stronger performance in the face of complex learning tasks.The matching accuracy shows that the transfer learning outperforms the vast majority of feature extraction methods.

关 键 词：迁移学习 模板保护 生物特征信息 特征提取 人脸识别 匹配精度 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP393.08[电子电信—信息与通信工程]