《计算机应用》唯一官方网站 ›› 2021, Vol. 41 ›› Issue (12): 3480-3484.DOI: 10.11772/j.issn.1001-9081.2021061107

• 第十八届中国机器学习会议(CCML 2021) • 上一篇    

优化三元组损失的深度距离度量学习方法

李子龙1,2(), 周勇2, 鲍蓉1, 王洪栋1   

  1. 1.徐州工程学院 信息工程学院,江苏 徐州 221018
    2.中国矿业大学 计算机科学与技术学院,江苏 徐州 221116
  • 收稿日期:2021-05-12 修回日期:2021-07-26 接受日期:2021-08-05 发布日期:2021-12-28 出版日期:2021-12-10
  • 通讯作者: 李子龙
  • 作者简介:周勇(1974—),男,江苏徐州人,教授,博士,CCF会员,主要研究方向:深度学习、计算机视觉
    鲍蓉(1968—),女,上海人,教授,博士,CCF会员,主要研究方向:深度学习、信息处理
    王洪栋(1986—),男,山东临沂人,讲师,博士,主要研究方向:图像处理、机器学习。
  • 基金资助:
    国家自然科学基金资助项目(61806206);江苏省建设系统科技项目(2018ZD077);徐州工程学院校级科研项目(XKY2019107);江苏省高校自然科学研究项目(20KJB170023)

Deep distance metric learning method based on optimized triplet loss

Zilong LI1,2(), Yong ZHOU2, Rong BAO1, Hongdong WANG1   

  1. 1.School of Information Engineering,Xuzhou University of Technology,Xuzhou Jiangsu 221018 China
    2.School of Computer Science and Technology,China University of Mining and Technology,Xuzhou Jiangsu 221116,China
  • Received:2021-05-12 Revised:2021-07-26 Accepted:2021-08-05 Online:2021-12-28 Published:2021-12-10
  • Contact: Zilong LI
  • About author:ZHOU Yong, born in 1974, Ph. D., professor. His research interests include deep learning, computer vision.
    BAO Rong, born in 1968, Ph. D., professor. Her research interests include deep learning, information processing.
    WANG Hongdong, born in 1986, Ph. D., lecturer. His research interests include image processing, machine learning.
  • Supported by:
    the National Natural Science Foundation of China(61806206);the Technology Project of Jiangsu Province Construction System(2018ZD077);the Scientific Research Project of Xuzhou University of Technology(XKY2019107);the Natural Science Research Project of Jiangsu Higher Education Institutions(20KJB170023)

摘要:

针对基于三元组损失的单一深度距离度量在多样化数据集环境下适应性差,且容易造成过拟合的问题,提出了一种优化三元组损失的深度距离度量学习方法。首先,对经过神经网络映射的三元组训练样本的相对距离进行阈值化处理,并使用线性分段函数作为相对距离的评价函数;然后,将评价函数作为一个弱分类器加入到Boosting算法中生成一个强分类器;最后,采用交替优化的方法来学习弱分类器和神经网络的参数。通过在图像检索任务中对各种深度距离度量学习方法进行评估,可以看到所提方法在CUB-200-2011、Cars-196和SOP数据集上的Recall@1值比之前最好的成绩分别提高了4.2、3.2和0.6。实验结果表明,所提方法的性能优于对比方法,同时在一定程度上避免了过拟合。

关键词: 深度距离度量, 深度学习, 三元组损失, 卷积神经网络, Boosting

Abstract:

Focused on the issues that the single deep distance metric based on triplet loss has poor adaptability to the diversified datasets and easily leads to overfitting, a deep distance metric learning method based on optimized triplet loss was proposed. Firstly, by thresholding the relative distance of triplet training samples mapped by neural network, and a piecewise linear function was used as the evaluation function of relative distance. Secondly, the evaluation function was added to the Boosting algorithm as a weak classifier to generate a strong classifier. Finally, an alternating optimization method was used to learn the parameters of the weak classifier and neural network. Through the evaluation of various deep distance metric learning methods in the image retrieval task, it can be seen that the Recall@1 of the proposed method is 4.2, 3.2 and 0.6 higher than that of the previous best score on CUB-200-2011, Cars-196 and SOP datasets respectively. Experimental results show that the proposed method outperforms the comparison methods, while avoiding overfitting to a certain extent.

Key words: deep distance metric, deep learning, triplet loss, Convolutional Neural Network (CNN), Boosting

中图分类号: