考虑难例挖掘和整体特征分布的损失函数设计  

Loss function designing considering hard example mining and global feature distribution

作　　者：吴皓[1] 王钰淏 田国会[1] 路飞[1] WU Hao;WANG Yuhao;TIAN Guohui;LU Fei(School of Control Science and Engineering,Shandong University,Jinan 250061,China)

机构地区：[1]山东大学控制科学与工程学院,山东济南250061

出　　处：《华中科技大学学报（自然科学版）》2021年第6期37-42,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61973192,U1813215,61973187,91748115,62773239);国家重点研发计划资助项目(2018YFB1307101)。

摘　　要：为了提高深度度量学习模型训练中的图像检索性能,在度量学习模型SoftTriple的基础上,对损失函数和网络结构进行改进.网络结构改进是在SoftTriple网络结构的基础上引入BNNeck模块;在损失函数设计中,首先添加难例挖掘函数对损失函数进行改进,然后通过使用高阶矩来表征整体特征分布的思路提出了新的损失函数.实验表明在度量学习数据集上的召回率和标准化互信息与改进前相比均有提高.与SoftTriple相比,添加难例挖掘函数实验最高将召回率提高了1.8%,标准化互信息值提高了2.2%;整体特征分布实验最高将召回率提高了0.9%,标准化互信息值提高了0.8%.In order to improve the performance of image retrieval in deep metric learning, new loss functions and new network structure were proposed based on the metric learning model Soft Triple. In the aspect of network structure improvement, BNNeck module was added to Soft Triple network structure.In the aspect of loss function design,two methods were proposed.Some loss functions were improved by adding the hard example mining functions. And a new loss function was proposed by using higher moments to characterize the global feature distribution. Some relative experiments were designed to verify our network and loss functions.The experimental results show that the recall rate and normalized mutual information are improved compared with those before improvement. Compared with Soft Triple, experiments of adding hard example mining functions improve the recall rate by1.8% and normalized mutual information by 2.2% at most.The global feature distribution experiments improve the recall rate by 0.9%and normalized mutual information by 0.8% at most.

关 键 词：度量学习 损失函数 难例挖掘 整体特征分布 高阶矩 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]