面向掌纹掌静脉识别网络轻量化的非对称双模态融合方法  

An asymmetric bimodal fusion method for lightweight palm print and palm vein recognition network

作　　者：林孙旗 徐家梦 郑瑜杰 王翀 王军 LIN Sunqi;XU Jiameng;ZHENG Yujie;WANG Chong;WANG Jun(Faculty of Electrical Engineering and Computer Science,Ningbo University,Ningbo 315211,China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211 [2]中国矿业大学信息与控制工程学院,江苏徐州221116

出　　处：《智能系统学报》2024年第5期1190-1198,共9页CAAI Transactions on Intelligent Systems

基　　金：科技部科技创新2030—“新一代人工智能”重大项目(2020AAA0107300);宁波市自然科学基金项目(20221JCGY010068);中国创新挑战赛(宁波)项目(2022T001).

摘　　要：深度学习已在掌纹掌静脉领域广泛应用,但随着任务使用场景的不断微型化、终端化,现有的深度学习模型往往难以在算力匮乏、内存有限的边缘设备上顺利部署。本文基于知识蒸馏方法提出了轻量化的掌纹掌静脉识别网络。根据模态特征提取复杂程度,为掌纹与掌静脉模态分别选用不同的网络深度。在常规知识蒸馏方法中引入新设计的模态特征损失函数,强化教师模型对各模态特征提取的指导作用。实验结果表明,该方法有效协调了模型大小与性能,为边缘计算环境下的生物特征识别技术提供了一种有效的解决方案。Deep learning has been widely used in palm print and palm vein recognition.However,with the continuous miniaturization and terminalization of task usage scenarios,it is often challenging to deploy current deep-learning models successfully on edge devices that suffer from limited computational power and memory constraints.In this study,we propose a lightweight palm print and palm vein recognition network based on knowledge distillation.First,we select different network depths for the palm print and palm vein modalities according to the complexity of their feature extraction.We introduce a novel modality feature loss function into the traditional knowledge distillation method to enhance the guiding role of the teacher model in the feature extraction of each modality.The experimental results demonstrate that this method effectively balances model size with performance and offers a viable solution for biometric recognition technologies within an edge computing environment.

关 键 词：深度学习 生物特征识别 掌纹掌脉识别 多模态网络 知识蒸馏 模型压缩 卷积神经网络 类激活图 

分 类 号：TP30[自动化与计算机技术—计算机系统结构]