摘要
针对近红外图像与人脸库里的可见光图像匹配的异质人脸识别(Heterogeneous face recognition,HFR),以及人脸识别所用到的深度卷积网络参数量和计算量大的问题,提出了用于学习近红外光图像和可见光图像之间跨光谱域不变的特征的轻量级人脸识别算法.首先使用大规模可见光图像数据集和改进的交叉熵损失函数,训练出一个类内紧凑、类间可分的可见光人脸识别模型,然后用改进的近红外光与可见光图像混合的三元组数据集配合的三重角度损失函数做迁移学习.相对于通用的深度学习方法,本方案对来自两种光谱域(近红外光和可见光)的人脸图像,都有很好的识别效果.同时通过研发轻量级的深度卷积网络,使算法可以在嵌入式设备中高效的运行,提高算法的工业应用价值.
Heterogeneous face recognition for matching near-infrared images with visible images in face database,as well as the problem of large depth convolution network parameters and computational complexity,a light weight face recognition algorithm for learning the invariant features between near-infrared light images and visible light images is proposed.Firstly,a large-scale visible light image dataset and improved cross-entropy loss function are used to train a compact,interclass visible light face recognition model.Then the improved triple Data Set of near infrared and visible images is used to do transfer learning with triple angle loss function.Compared with general deep learning methods,this scheme has a good recognition effect on face images from two spectral domains(nearinfrared and visible light).At the same time,by developing a lightweight deep convolution network,the algorithm can run efficiently in embedded devices and improve the industrial application value of the algorithm.
作者
张典
汪海涛
姜瑛
陈星
ZHANG Dian;WANG Hai-tao;JIANG Ying;CHEN Xing(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第4期807-811,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61462049)资助。
关键词
异质人脸识别
迁移学习
三重损失函数
轻量级深度卷积网络
heterogeneous face recognition
transfer learning
triple angle loss function
light weight deep convolution network
