测绘科学技术学报2025，Vol.41Issue(2)：163-172,10.DOI:10.3969/j.issn.1673-6338.2025.02.008
融合多尺度学习型特征与注意力机制的多源遥感图像匹配
Multi-source Remote Sensing Image Matching by Combining Multi-scale Deep Learning Feature and Attention Mechanism
摘要
Abstract
To tackle the challenge of local feature instability due to significant radiometric discrepancies in multi-source remote sensing imagery,which results in feature mismatch,an innovative matching approach are introduced in this paper.This approach leverages learning-based features that beyond local semantics and global positional in-formation,and integrates a Unet-like backbone network with a symmetric structure and multi-scale feature extrac-tion,along with a Transformer attention mechanism to ensure feature robustness.Firstly,the developed Unet-like backbone network is used to extract learning-based features to eliminate the positional instability of multi-scale lo-cal features.Then,the Transformer attention mechanism is employed to enhance the global representational capaci-ty of the features,and a consistency matching strategy is adopted to refine the precise matching points.The effec-tiveness and reliability of the proposed method in matching optical images with optical,infrared and SAR remote sensing images under multi-platform,multi-temporal,and multi-modal conditions have been verified on two public datasets and one dataset containing diverse scenes.This provides a reference for the registration of multi-source re-mote sensing images.关键词
多源遥感图像/图像匹配/学习型特征/注意力机制/深度学习Key words
multi-source remote sensing image/image matching/learning feature/attention mechanism/deep learning分类
测绘与仪器引用本文复制引用
张永显,薛源,徐梦珍,邢轩玮,马国锐,逄一哲..融合多尺度学习型特征与注意力机制的多源遥感图像匹配[J].测绘科学技术学报,2025,41(2):163-172,10.基金项目
水圈科学与水利工程全国重点实验室团队重点项目(sklhse-TD-2024-E01) （sklhse-TD-2024-E01）
国家自然科学基金重点项目(U2243222). （U2243222）
