摘要
FaceNet作为人脸识别的一大跨越,以其高精度、低硬件配置等优势被广泛应用于各个人脸识别相关领域。本文开源了首个餐厅支付场景下的中国人脸数据集CN-Face,该数据集拥有13000人的人脸图像,总计100000张。此外,本文以CA-SIA-WebFace作为训练集,利用改进后的MobileNet主干网络,采取不同的注意力机制添加方法,改变激活函数并且融入Contextual Transformer模块,大大降低了参数量和识别速度,显著提升了人脸识别精度。相较于原版FaceNet,在LFW测试集下,准确率达到98.79%,提升了2.74%,在CN-Face数据集中准确率达到95.22%,提升了1.35%。
FaceNet,as the major progress in face recognition,has been widely used in related face recognition fields due to its advantages of high precision and low hardware configuration requirements.The laboratory collects face data from multiple restaurants and combines them to form a Chinese face dataset(CN-Face)with more than 13000 face IDs and 100000 face images.Explore the use of CASIA-WebFace as the training set,use the improved MobileNet backbone network,adopt different attention mechanism addition methods,change the activation function and integrate the Contextual Transformer module,which greatly reduces the number of parameters and recognition speed,and significantly improves face recognition precision.Compared with the original FaceNet,under the LFW test set,the accuracy rate reaches 98.79%,an increase of 2.74%.At the same time,the accuracy rate reaches 95.22%under the Chinese face dataset(CN-Face)collected in the laboratory,an increase of 1.35%.
作者
陈经纬
熊继平
程汉权
CHEN Jingwei;XIONG Jiping;CHENG Hanquan(College of Physics and Electronic Information Engineering,Zhejiang Normal University,Jinhua 321004,Zhejiang,China)
出处
《智能计算机与应用》
2024年第3期61-66,共6页
Intelligent Computer and Applications
基金
金华市公益项目(2021-4-116)。
