一种基于排序代理锚损失的深度度量学习算法  

Ranked Proxy Anchor Loss Based Deep Metric Learning Algorithm

作　　者：张兵 陈海燕[1,3] 侯夏晔 袁立罡 刘振亚[1,3] ZHANG Bing;CHEN Hai-yan;HOU Xia-ye;YUAN Li-gang;LIU Zhen-ya(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;College of Civil Aviation,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Collaborative Innovation Center of Novel Software Technology and Industrialization,Nanjing 210023,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106 [2]南京航空航天大学民航学院,南京211106 [3]软件新技术与产业化协同创新中心,南京210023

出　　处：《小型微型计算机系统》2022年第10期2035-2039,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61501229)资助;中央高校基本科研业务费专项资金项目(NS2019054,NS2020045)资助.

摘　　要：深度度量学习根据特定的度量损失函数对神经网络进行不同的训练,直接学习原始图像空间到语义特征嵌入空间的非线性投影.针对现有代理锚损失函数在不考虑数据分布的情况下,将同一类正样本压缩到嵌入空间中的某个代理锚点,而造成同类样本相似结构丢失的问题,本文提出了一种新的基于排序驱动策略的代理锚损失函数.该损失函数使用了排序列表损失中对样本对进行排序的思想,将正样本排在负样本之前,并通过约束正样本对的距离小于阈值来尽可能地保留同类内部的相似结构.最后,在两个标准数据集上的实验证明了本文所提方法在图像分类问题上的有效性和优越性.Deep metric learning trains neural networks differently according to a specific metric loss function,and directly learns the nonlinear projection from original image space to semantic feature embedding space.This paper proposes a new proxy anchor loss based on ranking-motivated strategy,aiming at the problem that the existing proxy anchor loss function compresses the same class of positive samples to a certain proxy anchor point in the embedding space,without considering the data distribution,resulting in the loss of the similar structure of the similar samples.The loss function uses the idea of sorting sample pairs in the ranked list loss for reference,ranks positive samples before negative samples,and constrains the distance between positive sample pairs to be less than a threshold to keep the similar structures within the same class as much as possible.Finally,experiments on two standard datasets verify the effectiveness and superiority of the proposed method on the image classification problem.

关 键 词：度量学习 深度学习 度量损失函数 代理锚损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]