基于集成对比学习的高分辨遥感图像搜索模型  

Hi-Res Remote Image Searching Based on Ensemble Contrastive Learning

作　　者：顾程成 孙勇[1,2] 程千禧 谭文安 Gu Chengcheng;Sun Yong;Cheng Qianxi;Tan Wenan(School of Computer and Information Engineering,Shanghai Polytechnic University,Shanghai 201209,China;Key Laboratory of Physical Geographic Environment of Anhui Province,Chuzhou University,Chuzhou 239000,China;School of Resources and Environmental Engineering,Anhui University,Hefei 230009,China)

机构地区：[1]上海第二工业大学计算机与信息工程学院,上海201209 [2]滁州学院实景地理环境安徽省重点实验室,安徽滁州239000 [3]安徽大学资源环境学院,安徽合肥230009

出　　处：《南京师范大学学报(工程技术版)》2025年第1期22-29,共8页Journal of Nanjing Normal University(Engineering and Technology Edition)

基　　金：安徽省教育厅重大重点科学研究项目(2022AH051113);安徽省重点实验室开放基金资助项目(2022PGE003);滁州学院科学研究基金重点项目(2022XJZD06).

摘　　要：为解决单一对比学习模型只关注图像的局部特征问题,提出一种基于集成对比学习的高分辨率遥感图像搜索模型.首先,采用动量更新无监督视觉表示学习与采样最近邻对比学习模型分别提取遥感图像的局部和全局特征,以学习出更好的遥感图像视觉表征.在此基础上,提出面向高分辨率遥感图像表征的集成对比学习模型,根据其在遥感图像分类任务的表现,对不同的对比学习器自适应地赋予不同的集成权重,通过统计对比学习特征预测准确率,进一步优化对比学习器的学习速度.最后,将集成对比学习模型运用于高分辨遥感图像搜索.在EuroSat、UCmerced、WHU-RS19、PatternNet等公开遥感图像数据集上的实验结果表明,所提出的模型在图像搜索任务中相较传统的对比方法有较稳定的准确率提升.To address the issue of currently many single contrastive learning models focusing only on local features of images,a new high resolution remote image searching model based on ensemble contrastive learning is proposed.Firstly,local and global features and learned from remote images by momentum contrastive learning and nearest neighbor contrastive learning methods.Secondly,on this basis,high resolution remote image searching model based on ensemble contrastive learning is proposed,which sets different weights for different contrastive learners according to their own performance in downstream classification tasks,and then further optimizes the learning speeds of learners with the accuracy of contrastive learning features.Finally,ensemble contrastive learning model is applied in Hi-Res remote image searching.Experiments are conducted by using public remote sensing image datasets including EuroSat,UCmerced,WHU-RS19,and PatternNet.The results indicate that the proposed model demonstrates a more stable improvement in accuracy compared with traditional contrastive methods for image searching tasks.

关 键 词：对比学习 集成学习 最近邻对比学习 动量更新对比学习 遥感图像搜索 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP18[自动化与计算机技术—控制科学与工程]