期刊文献+

基于多任务级联CNN与中心损失的人脸识别

Face Recognition Based on Multi-Task Cascade Convolution CNN and Center Loss Function
在线阅读 下载PDF
收藏 分享 导出
摘要 开源机器学习库DLIB中的人脸检测对齐任务的运行时间长,检测精度不高,以及用三元组损失训练的模型收敛速度较慢。针对以上三点不足,提出了用mtcnn代替DLIB做人脸检测,并且以用softmax loss和center loss相结合的总损失函数来训练卷积神经网络。首先,将公开的海量人脸数据集做人脸对齐。然后,以总损失函数作为监督信号来完成BP前向传播,使得类内距离小,类间距离大,提高模型的特征辨识能力。最后,对人脸特征进行embedding,由高维度映射到低维度,减少参数量,提高识别率。实验表明,人脸检测对齐后的测试集比原始测试集在识别率上要高1%左右,并且mtcnn在运算速度和检测精度上优越于DLIB。用总损失函数训练的模型经过调参调优阶段在LFW标准测试集上识别率为99.6%,同时在megface标准的人脸库上也有较高的识别率。用自己创建的三张人脸图片成功验证了度量学习的特性。 For the open source machine learning library(DLIB),the task of face alignment has long running time and low detection precision,as well as slow convergence rate of the model trained with triple loss.In order to solve the above three problems,this paper used mtcnn instead of DLIB to detect human face,also the total loss function combined with softmax loss and center loss was used to train convolutional neural network.First of all,this method aligned a human face with an open mass of face datasets.Afterward,the BP forward propagation was completed by using the total loss function as the supervisory signal,which makes the intra-class distance small and the inter-class distance large,as well as improves the feature identification ability of the model.Finally,the facial features were embedded and mapped from high dimension to low dimension so as to reduce the amount of parameters,at the same time,to improve the recognition rate.Experiments show that the face detection alignment test set is about 1%higher than the original test set in recognition rate.In addition,mtcnn is superior to DLIB in operation speed and detection accuracy.The model trained with the total loss function has a recognition rate of 99.6%on the LFW standard test set after tuning parameters and a high recognition rate on the face database of the megface standard.Three face images created by ourselves were successfully used to verify the characteristics of metric learning.
作者 王灵珍 赖惠成 WANG Ling-zhen;LAI Hui-cheng(School of Information Science and Engineering,Xinjiang University,Urumqi Xinjiang 830046,China)
出处 《计算机仿真》 北大核心 2020年第8期398-403,共6页 Computer Simulation
关键词 人脸识别 中心损失 度量学习 卷积神经网络 人脸对齐 人脸检测 Face recognition Center loss Metric learning Convolutional neural networks(CNN) Face alignment Face detection
  • 相关文献

参考文献5

二级参考文献21

共引文献64

论文智能改写系统
维普数据出版直通车
投稿分析
职称考试

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部 意见反馈