基于特征融合与局部对比学习的图像检索  

Image retrieval based on feature fusion and local contrastive learning

作　　者：何强 张卫华[2] 周激流[1] He Qiang;Zhang Weihua;Zhou Jiliu(School of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China;School of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]四川大学计算机学院,成都610065

出　　处：《现代计算机》2024年第23期42-47,共6页Modern Computer

基　　金：四川省科技计划项目(2022YFQ0047)。

摘　　要：图像检索的目标是从庞大的图像数据库中找出与查询图像最相似的若干图像。近年来,使用全局特征检索和局部特征重排序的双阶段图像检索方法取得了出色的性能表现,但重排序阶段的加入导致了整体检索响应时间慢的问题,而且局部特征主要通过全局的分类损失函数进行训练,使局部特征中产生了大量的冗余信息。针对这些问题,提出一种基于特征融合和局部对比学习的单阶段图像检索方法。在图像检索任务中的实验证明,本文提出的图像检索方法表现出良好的性能,为图像检索的研究和应用提供了有益的启示。Image retrieval is aim at finding out several images that are most similar to the query image from the huge image database.In recent years,the two‑stage image retrieval methods using global feature retrieval and local feature reranking have achieved excellent performance,but the reranking stage leads to the problem of slow response time,and local features are trained mainly by the global classification loss function,this method results in a large amount of redundant information in local features.To solve this problem,this paper introduces a single‑stage image retrieval method based on feature fusion and local contrastive learning.The experimental results in image retrieval task show that the image retrieval method proposed in this paper demonstrates strong performance,and provides useful enlightenment for the research and application of image retrieval.

关 键 词：图像检索 特征融合 对比学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]