摘要
在开集协议下设计了一种基于角度距离损失函数和密集连接卷积神经网络的人脸识别算法,以实现深度人脸识别。所设计的网络结构使用基于角度距离的损失函数,让人脸特征的区分度更高,符合特征的理想分类标准。同时,所提出的神经网络结构采用先进的密集连接模块,在很大程度上减少了传统网络结构的参数冗余。经过大量的分析和实验,该算法在LFW数据集上的人脸识别准确率达到了99.45%,在MegaFace数据集上的人脸确认任务和人脸验证任务中的人脸识别准确率分别为72.534%和85.348%,因此所提算法在人脸识别任务中具有较高的优越性。
A face recognition algorithm based on the angular distance loss function and densely connected convolutional neural network is proposed under the open-set protocol to achieve deep face recognition.The loss function based on angle distance is adopted in the proposed network structure,which makes the facial features more distinguishable and meets the ideal criteria of feature classification.At the same time,the advanced dense connection module is adopted in the proposed neural network structure,which greatly reduces the parameter redundancy of the traditional network structure.After extensive analysis and repeated experiments,the face recognition accuracy reaches 99.45%on the LFW dataset,and the recognition accuracy rates of face identification task and face verification task on MegaFace dataset are 72.534% and 85.34%,respectively.The superiority of the proposed algorithm is confirmed in the face recognition domain.
作者
龙鑫
苏寒松
刘高华
陈震宇
Long Xin;Su Hansong;Liu Gaohua;Chen Zhenyu(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2018年第12期402-413,共12页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61471260)
关键词
机器视觉
人脸识别
卷积神经网络
深度学习
角度损失函数
密集连接
machine vision
face recognition
convolutional neural network
deep learning
angular loss function
dense connection
