基于差异特征补偿Transformer的遥感图像变化检测  

Differential Feature Compensation Transformer for Remote Sensing Image Change Detection

作　　者：孙芸倩 何强 陈琳琳 SUN Yunqian;HE Qiang;CHEN Linlin(School of Science,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Beijing University of Civil Engineering and Architecture,Institute of Big Data Modeling Theory and Technology,Beijing 102616,China)

机构地区：[1]北京建筑大学理学院,北京102616 [2]北京建筑大学大数据建模理论与技术研究所,北京102616

出　　处：《河北师范大学学报（自然科学版）》2024年第6期541-551,共11页Journal of Hebei Normal University：Natural Science

基　　金：国家自然科学基金(62072024,12301581);北京市教育委员会科学研究计划项目(KM202110016001,KM202210016002)。

摘　　要：遥感图像变化检测(remote sensing image change detection.RSICD)的目标是检测双时图像之间发生变化的区域.为了准确获取感兴趣的差异特征图,本文提出了差异特征补偿Transformer(differential feature compensation Transformer,DFCT)模型来解决RSICD问题,设计了差异特征补偿分支来有效地学习差异特征.为了更好地连接高级和低级特征,引入了特征融合模块.具体而言,首先使用共享骨干网络来提取给定图像对的特征图.然后,将共享骨干网络输出的特征映射为新的表征集,连接起来输入到Transformer编码器以获得每个时态的细粒度特征图.最后特征融合模块利用差异特征补偿分支补充双增强特征的粗粒度特征信息,将该特征与分支得到的差异补偿特征进行融合.最后,通过预测头获得精确的变化图.该方法在3个数据集上的F1分数分别达到了93.46%,92.25%和96.27%,交并比分别达到了87.71%,85.62%和92.82%,与其余10种当前主流方法相比均为最优,实验结果证明了该方法的有效性.The objective of remote sensing image change detection(RSICD) is to identify changes between bitemporal images.This paper proposes a differential feature compensation Transformer(DFCT)model to address the RSICD problem and designs a new feature compensation branch to learn different features.Additionally,a feature fusion module is introduced to better integrate high-level and low-level features.Specifically,a shared backbone network is utilized to extract the feature maps of an image pair,which are then fed into the Transformer encoder to obtain fine-grained feature maps for each temporal.Finally,an accurate change map is obtained using the prediction head.The F1 scores of the proposed method on three datasets achieved 93.46 %,92.25 % and 96.27 %,and the IoU achieved 87.71 %,85.62 % and 92.82 %,respectively,which outperformed the other 10 mainstream methods.The experimental results confirmed the effectiveness of the proposed method.

关 键 词：遥感 变化检测 注意力机制 差异特征补偿 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]