作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (3): 229-236. doi: 10.19678/j.issn.1000-3428.0054327

• 图形图像处理 • 上一篇    下一篇

基于多任务学习的人脸属性识别方法

李亚, 张雨楠, 彭程, 杨俊钦, 刘淼   

  1. 广州大学 计算机科学与网络工程学院, 广州 510006
  • 收稿日期:2019-03-21 修回日期:2019-05-09 发布日期:2019-05-20
  • 作者简介:李亚(1980-),女,副教授、博士,主研方向为计算机视觉、模式识别;张雨楠、彭程、杨俊钦,本科生;刘淼,副教授、博士。
  • 基金资助:
    国家大学生创新训练项目(201711078017,201811078025);广州市教育局市属高校科研项目(1201620302)。

Face Attributes Recognition Method Based on Multi-Task Learning

LI Ya, ZHANG Yu'nan, PENG Cheng, YANG Junqin, LIU Miao   

  1. School of Computer Science and Cyber Engineering, Guangzhou University, Guangzhou 510006, China
  • Received:2019-03-21 Revised:2019-05-09 Published:2019-05-20

摘要: 针对传统深度卷积神经网络模型复杂、识别速度慢的问题,提出一种基于多任务学习的人脸属性识别方法。通过轻量化残差模块构建基础网络,根据属性类之间的关联关系设计共享分支网络,以大幅减少网络参数和计算开销。以多任务学习的方式联合优化各分支网络与基础网络的参数,利用关联属性间的共同特征实现人脸属性识别。采用带权重的交叉熵作为损失函数监督训练网络模型,改善正负样本数不均衡问题。在公开数据集CelebA上的实验结果表明,该方法的识别错误率低至8.45%,空间开销仅2.7 MB,在CPU上每幅图预测时间低至15 ms,方便部署在资源有限的移动或便携式设备上,具有实际应用价值。

关键词: 人脸属性识别, 轻量化残差模块, 深度卷积神经网络, 模型压缩, 多任务学习

Abstract: To address the problem of complex model and slow recognition speed of traditional deep convolutional neural network,this paper proposes a face attributes recognition method based on multi-task learning.The underlying network is constructed by the lightweight residual modules.According to the correlations between attribute classes,the sharing branch networks are designed to largely reduce the network parameters and calculation costs.Then the parameters of the branch networks and underlying network are jointly optimized in the manner of multi-task learning.The shared features of correlated attributes are used to achieve better recognition effect.The weighted cross entropy is taken as the supervised training network model of the loss function,so as to improve the disequilibrium of positive and negative samples.Experimental results on the public dataset CelebA show that the recognition error rate of the proposed algorithm can be reduced to 8.45% and the space cost is only 2.7 MB.Besides,the prediction time of each image on CPU is reduced to 15 ms,which is suitable for the resource-limited portable devices and valuable in real life applications.

Key words: face attributes recognition, lightweight residual module, deep convolutional neural network, model compression, multi-task learning

中图分类号: