基于分数阶微分的卷积神经网络的人脸识别
作者:
作者单位:

四川大学计算机学院

作者简介:

通讯作者:

中图分类号:

TP391.4

基金项目:

国家重点研究与发展计划基金(2018YFC0830300); 国家自然科学基金(61571312)


Convolution neural network face recognition based on fractional differential
Author:
Affiliation:

College of Computer Science, Sichuan University

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    近年来,人脸识别技术作为一种用来抓取生物面部特征信息以及匹配现有数据库中人脸数据的有力手段,以其无接触性、可远距离实施等优点在越来越多的场景中得到了应用. 针对在自然无约束条件下,受到光照、姿势和背景环境等因素的影响,设备捕捉到的人脸图像在现有的人脸识别模型中识别率依然不足的情况,本文提出了一种基于分数阶微分改进的残差网络(ResNet)人脸识别方法. 本方法通过在原有网络模型结构中增加注意力机制来增强人脸特征提取,融合不同通道和空间的信息提升网络的健壮性,同时利用分数阶微分对节点函数进行处理,增加卷积块提取更多的人脸细节信息,最后使用ArcFace损失函数来优化模型,在网络中进行迭代训练完成人脸识别. 实验结果显示:改进后的网络模型在现有的一些数据集(如LFW、AgeDB-30、CFP-FP等)上表现出更好的识别性能和更强的鲁棒性.

    Abstract:

    In recent years, face recognition technology as a powerful means to capture biological facial feature information and match face data in existing databases has been applied in more and more scenes with its advantages of non-contact and remote implementation. Due to the influence of illumination, posture, background environment and other factors under the natural unconstrained condition, the recognition rate of the face images captured by the devices is still insufficient in the existing face recognition model. This paper proposes a face recognition method based on fractional differential improved residual network (ResNet), by adding attention mechanism to the original network model structure. At the same time, fractional differential is used to process node functions, and convolution blocks are added to extract more face details. Finally, Arcface loss function is used to optimize the model, and iterative training is carried out in the network to complete face recognition. The experimental results show that the proposed model has better recognition performance and stronger robustness on the existing data sets such as LFW, AgeDB-30, CFP-FP.

    参考文献
    相似文献
    引证文献
引用本文

引用本文格式: 彭朝霞,蒲亦非. 基于分数阶微分的卷积神经网络的人脸识别[J]. 四川大学学报: 自然科学版, 2022, 59: 012001.

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-01-04
  • 最后修改日期:2021-06-19
  • 录用日期:2021-07-16
  • 在线发布日期: 2022-01-19
  • 出版日期: