基于上下文感知与多尺度注意力的遥感变化检测  

Remote Sensing Change Detection Based on Context Perception and Multi-Scale Attention

作　　者：董晨 郑禄[1,3] 于舒 饶白云 DONG Chen;ZHENG Lu;YU Shu;RAO Baiyun(School of Computer Science,South-Central Minzu University;Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprise;Hubei Provincial Engineering Research Center of Agricultural Blockchain and Intelligent Management,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院 [2]湖北省制造企业智能管理工程技术研究中心 [3]农业区块链与智能管理湖北省工程研究中心,湖北武汉430074

出　　处：《软件导刊》2023年第11期65-70,共6页Software Guide

基　　金：国家民委中青年英才培养计划项目(MZR20007);中央高校基本科研业务费专项资金资助项目(CXY22008);湖北省科技重大专项(2020AEA011);新疆维吾尔自治区区域协同创新专项(2022E02035)。

摘　　要：遥感图像中存在大小不一的地物,因此具有多尺度特性。现有变化检测方法大多采用多尺度分割以获得更完整、丰富的局部特征,但忽略了非局部位置特征(全局上下文信息)的重要性,导致提取特征差异化表示较弱。为了解决该问题,提出一个基于上下文感知与多尺度注意力的孪生网络SPAN。在特征提取模块,使用孪生金字塔结构提取不同尺度的特征进行融合以获取鲁棒性更强的特征;在注意力增强模块,引入一个结合全局上下文的金字塔分割注意力模块PSG,以进一步提取目标特征并增强差异化表示。实验表明,所提模型相较于现有模型,在变化检测标准数据集CDD和LEVIR-CD上的F1分别达到0.944、0.812。消融实验表明,PSG相较于传统的多尺度注意力,能充分提取具有判别性的特征,证明了全局上下文信息对变化检测任务的重要性。Remote sensing images have multi-scale characteristics due to the presence of objects of varying sizes.Most existing change detection methods use multi-scale segmentation to obtain more complete and rich local features,but ignore the importance of non local position features(global context information),resulting in weak differential representation of extracted features.To address this issue,a twin network SPAN based on context awareness and multi-scale attention is proposed.In the feature extraction module,twin pyramid structures are used to extract features of different scales for fusion to obtain more robust features;In the attention enhancement module,an attention module PSG that combines global context pyramid segmentation is introduced to further extract target features and enhance differential representation.The experiment shows that compared to existing models,the F1 of the proposed model on the change detection standard dataset CDD and LEVIR-CD reaches 0.944 and 0.812,respectively.The ablation experiment shows that PSG can fully extract discriminative features compared to traditional multi-scale attention,demonstrating the importance of global contextual information for change detection tasks.

关 键 词：变化检测 孪生 多尺度 注意力 全局上下文 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]