计算机工程2025，Vol.51Issue(5)：305-313,9.DOI:10.19678/j.issn.1000-3428.0068894
基于类中心优化辅助三元组损失的遥感图像检索
Remote Sensing Image Retrieval Based on Class Center Optimization Added Triplet Loss
摘要
Abstract
The key to remote sensing image retrieval is to efficiently and accurately retrieve target samples from massive images.Intraclass samples in remote sensing images are dispersed and exhibit large variance.Traditional remote sensing image retrieval based on limited samples cannot effectively learn the differences between intraclass samples.The existing Cross-Batch Memory(XBM)method has triplet pairing redundancy and complex computations.A remote sensing image retrieval method based on Class Center Optimization added for Triplet Loss(CCO-TL)is proposed to address these problems.CCO-TL uses class center features to limit the distance between positive samples within a class,assisting in optimizing the triplet loss and achieving interclass separation.Simultaneously,samples within a class are clustered and compacted,generating optimized sample features.By improving the XBM module,a Batch Feature Queue(BFQ)is obtained to store the feature vectors of previous training batches,and by changing the triplet pairing method,sample information is mined fully,data redundancy problems are solved,and the training time is reduced.Simultaneously,the BFQ module is used for the real-time calculation of class center point features,replacing the estimated values of traditional methods with calculated values.Experimental results show that the network model trained with the triplet loss function based on real class center feature assisted optimization has a stronger learning ability between samples,more intraclass clustering,and more obvious interclass differentiation.The proposed method is evaluated in terms of the Recall@K metric on four remote sensing datasets.The proposed method achieves accuracies of 93.1％,87.2％,97.1％,and 82.2％,on the UCMD,AID,PN,and OP datasets,respectively,outperforming other methods.关键词
图像检索/深度度量学习/三元组损失/类中心/批次Key words
image retrieval/deep metric learning/triplet loss/class center/batch分类
信息技术与安全科学引用本文复制引用
郑宗生,霍志俊,高萌,王政翰,周文睆,张月维..基于类中心优化辅助三元组损失的遥感图像检索[J].计算机工程,2025,51(5):305-313,9.基金项目
国家自然科学基金(41671431) （41671431）
上海市科委地方能力建设项目(19050502100) （19050502100）
广州气象卫星地面站项目(D-8006-23-0157). （D-8006-23-0157）
