[1]杨孟渭,张索非,吴晓富,等.基于多尺度三元组损失的层级图像检索算法[J].计算机技术与发展,2025,(04):80-85.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0380]
　YANG Meng-wei,ZHANG Suo-fei,WU Xiao-fu,et al.Hierarchical Image Retrieval with Multi-scale Triplet Loss[J].,2025,(04):80-85.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0380]
点击复制

基于多尺度三元组损失的层级图像检索算法()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2025年04期
页码:
80-85
栏目:
人工智能
出版日期:
2025-04-10

文章信息/Info

Title:
Hierarchical Image Retrieval with Multi-scale Triplet Loss
文章编号:
1673-629X(2025)04-0080-06
作者:
杨孟渭; 张索非; 吴晓富; 周全
南京邮电大学,江苏 南京 210003
Author(s):
YANG Meng-wei; ZHANG Suo-fei; WU Xiao-fu; ZHOU Quan
Nanjing University of Posts and Telecommunications,Nanjing 210003,China
关键词:
度量学习; 多语义尺度; 层级图像检索; 对比学习; 深度学习
Keywords:
metric learning; multiple semantic scales; hierarchical image retrieval; contrastive learning; deep learning
分类号:
TP31
DOI:
10.20165/j.cnki.ISSN1673-629X.2024.0380
摘要:
现有大多数度量学习方法的目标是扩大类间样本的距离,同时减小类内样本的距离,在多层语义场景中,此类度量学习方法是不太适用的。 例如,金丝猴属于猴子,与猫不相似,但从哺乳动物角度看,金丝猴和猫又属于同一类。 因此,多层语义场景下存在的主要问题表现为:在较小的语义尺度上的负对可以是在较大语义尺度上的正对,因此拉开较小语义尺度上的负对距离会损害较大语义尺度上的类结构,反之亦然。 为此,该文提出了一种新的多尺度三元组损失(Multi-Scale Triplet Loss,MSTL)来解决层级图像检索问题。 为了使得每个类别的嵌入空间更加紧凑,只选取每个语义尺度标签都不相同的类别作为负样本,同时提出了更适用于细粒度语义尺度的一种新的代理损失。 最后,通过在已有模型上赋予 MSTL 适合的权重,以进一步提升模型的鲁棒性与检索性能。 在三个 DyML 数据集上进行的大量实验表明,该方法优于现有流行方法。
Abstract:
Most existing metric learning methods aim to expand the distance of inter-class samples while reducing the distance of intra-class samples,and such metric learning methods are less applicable at multiple semantic scales. For example,golden monkeys belong to monkeys,which are not similar to cats, but from the perspective of mammals, golden monkeys and cats belong to the same class.Therefore,the main problem existing in multiple semantic scales manifests itself in the fact that a negative pair at a finer semantic scale can be a positive pair at a coarser semantic scale, and thus distancing negative pairs at finer semantic scales compromises the class structure at coarser semantic scales,and vice versa. For this reason,we propose a new Multi-Scale Triplet Loss (MSTL) to solve the hi-erarchical image retrieval problem. In order to make the embedding space of each class more compact,only the classes with different labels at each semantic scale are selected as negative samples,and a new proxy loss that is more applicable to fine semantic scales is pro-posed. Finally,the robustness and retrieval performance of the model is further improved by assigning suitable weights to MSTL on the existing model. Extensive experiments on three DyML datasets show that the proposed method outperforms existing popular methods.

相似文献/References:

[1]史作婷,吴 迪,荆晓远,等.类不平衡稀疏重构度量学习软件缺陷预测[J].计算机技术与发展,2018,28(06):125.[doi:10.3969/ j. issn.1673-629X.2018.06.028]
　SHI Zuo-ting,WU Di,JING Xiao-yuan,et al.Prediction of Defect of Class-imbalance Sparse Reconstruction Metric Learning Software[J].,2018,28(04):125.[doi:10.3969/ j. issn.1673-629X.2018.06.028]
[2]王晴[],荆晓远[][],朱阳平[],等. 基于局部稀疏重构度量学习的软件缺陷预测[J].计算机技术与发展,2016,26(11):54.
　WANG Qing[],JING Xiao-yuan[][],ZHU Yang-ping[],et al. Software Defect Prediction of Metric Learning Based on Local Sparse Reconstruction[J].,2016,26(04):54.
[3]刘 军,孙 庆,刘 玮,等.基于 U-Net 网络和椭圆度量学习的防震锤锈蚀识别[J].计算机技术与发展,2020,30(11):163.[doi:10. 3969 / j. issn. 1673-629X. 2020. 11. 030]
　LIU Jun,SUN Qing,LIU Wei,et al.Identification of Anti-vibration Hammer Corrosion of High-voltage Transmission Lines Based on U-Net Network and Elliptic Metric Learning[J].,2020,30(04):163.[doi:10. 3969 / j. issn. 1673-629X. 2020. 11. 030]
[4]刘 东,胡峻林*.基于度量学习的步态识别比较研究[J].计算机技术与发展,2022,32(09):89.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 014]
　LIU Dong,HU Jun-lin *.Comparative Study of Gait Recognition Based on Metric Learning[J].,2022,32(04):89.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 014]
[5]江雨燕,吕 魏,李 平,等.基于无监督的非对称度量学习优化行人再识别[J].计算机技术与发展,2022,32(09):126.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 020]
　JIANG Yu-yan,LYU Wei,LI Ping,et al.Unsupervised Asymmetric Metric Learning for Person Re-identification Optimization[J].,2022,32(04):126.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 020]
[6]行阳阳,张索非,宋 越,等.一种面向商品检索的多尺度度量学习方法[J].计算机技术与发展,2024,34(01):65.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 010]
　XING Yang-yang,ZHANG Suo-fei,SONG Yue,et al.A Multi-scale Metric Learning Approach for Product Retrieval[J].,2024,34(04):65.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 010]

更新日期/Last Update: 2025-04-10