基于交叉量化和样本校正的自监督遥感图像检索  

Self-supervised Remote Sensing Image Retrieval Based on Cross-quantization and Sample Correction

作　　者：潘丽丽[1] 瞿栋梁 尹晶晶 马雪强 PAN Lili;QU Dongliang;YIN Jingjing;MA Xueqiang(School of Computer Science and Mathematics,Central South University of Forestry and Technology,Changsha 410000,China;School of Electronic Information and Physics,Central South University of Forestry and Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学计算机与数学学院,湖南长沙410000 [2]中南林业科技大学电子信息与物理学院,湖南长沙410000

出　　处：《郑州大学学报(工学版)》2025年第2期60-66,共7页Journal of Zhengzhou University(Engineering Science)

基　　金：湖南省自然科学基金资助项目(2021JJ31164);湖南省教育厅科学研究重点项目(22A0195)

摘　　要：自监督遥感图像检索模型由于标签缺失导致使用错误的样本对进行学习,从而产生抽样偏差,影响图像表征的准确性,为此提出一种基于交叉量化和样本校正的自监督遥感图像检索模型(CQSC)。首先,为了降低数据存储和处理负载,将传统对比学习中的映射层与乘积量化相结合,压缩高维图像数据,提高了检索效率;其次,使用交叉学习策略,最大化检索模型中特征映射前后的交叉相似性,增强模型的特征生成能力和检索精度;最后,针对自监督遥感图像检索中因标签缺失导致的抽样偏差,设计自适应纠错标签,标注训练样本,校正训练过程中存在的错误负样本。在UCMerced和EuroSAT数据集上进行了大量实验,结果表明:与PLSH方法相比,在UCMerced数据集上,所提方法mAP@20平均提升了2.52百分点;在EuroSAT数据集上取64 bits时,所提方法mAP@100提升了3.83百分点。In view of the fact that the self-supervised remote sensing image retrieval model by using the sample pair for learning due to the lack of labels,resulted in sampling bias and affecting the accuracy of image representation,a self-supervised remote sensing image retrieval model based on cross-quantization and sample correction(CQSC)was proposed.Firstly,in order to reduce the load of data storage and processing,the mapping layer and product quantization in traditional contrast learning were combined to compress the high-dimensional image data and improve the retrieval efficiency.Secondly,the cross-learning strategy was used to maximize the cross-similarity before and after the feature mapping in the retrieval model,and the feature generation ability and retrieval accuracy of the model were enhanced.Finally,design adaptive correction labels to annotate training samples,correct erroneous negative samples during training and address sampling bias caused by missing labels in self-supervised remote sensing image retrieval.Experiments on UCMerced and EuroSAT datasets showed that compared with PLSH,mAP@20 of CQSC improved by 2.52 percentage points on average on UCMerced,and mAP@100 of CQSC improved by 3.83 percentage points on EuroSAT with 64bits.

关 键 词：遥感图像检索 对比学习 乘积量化 交叉学习 

分 类 号：P407.8[天文地球—大气科学及气象学] TP183[自动化与计算机技术—控制理论与控制工程]