基于局部-全局语义特征增强的遥感影像变化检测网络模型  

Local-global Semantic Feature Enhancement Model for Remote Sensing Imagery Change Detection

作　　者：高建文 管海燕 彭代锋 许正森 康健 季雅婷 翟若雪 GAO Jianwen;GUAN Haiyan;PENG Daifeng;XU Zhengsen;KANG Jian;JI Yating;ZHAI Ruoxue(School of Remote Sensing and Surveying Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,南京210044 [2]南京信息工程大学电子与信息工程学院,南京210044

出　　处：《地球信息科学学报》2023年第3期625-637,共13页Journal of Geo-information Science

基　　金：国家自然科学基金项目(41971414、41801386);江苏省2022年省级大学生创新创业训练计划(1514072201418)。

摘　　要：尽管卷积神经网络(Convolutional Neural Network,CNN)已广泛应用于遥感影像变化检测任务,但CNN感受野有限,难以有效提取全局语义信息。针对上述问题,本文提出一种用于变化检测任务的端到端局部-全局特征增强的编-解码网络模型(Local-Global Feature Enhancement Network,LGE-Net)。在编码部分,LGE-Net利用CNN_ResNet34骨干网络分别获得双时相遥感影像局部语义特征,级联后输入Locally Enhanced Transformer(LE-Transformer)结构捕获远距离语义依赖,提取深层全局语义特征。在解码部分,嵌入语义增强模块(Context Enhancement Module,CEM)连接解码特征与多尺度局部特征,实现变化对象的准确定位与分割。此外,针对LE-Transformer各分块内部以及相邻分块序列间缺乏局部信息交互问题,设计了局部特征增强前馈网络(Locally Enhanced Feed Forward Network,LEFFN)。在LEVIR-CD和CDD变化检测数据集上的综合对比试验表明,本文提出LGE-Net模型取得的F1-score分别达到91.06%和94.78%,显著优于其他对比模型,可更加精准识别变化区域,进一步减少误检以及漏检率,且具有良好的泛化能力。Convolutional Neural Network(CNN)has achieved promising results in change detection using remote sensing images.However,CNN performs poorly on global semantic information extraction due to its limited receptive field.To this end,we propose an end-to-end encoding-decoding local-global feature enhancement network,termed as LGE-Net,which introduces locally enhanced Transformers(LE-Transformer)for capturing global semantic feature representation.Specifically,the LGE-Net uses the CNN backbone network to obtain local semantic features of dual-phase remote sensing images and cascades the extracted local features into the LE-Transformer layer to extract deep global semantic features.Then,in the decoder,the features are cascaded,up-sampled,and finally connected with multi-scale local features by semantic enhancement modules(CEMs).In addition,a local feature-enhanced feed-forward network(LEFFN)is designed to enhance local information interaction in the LE-Transformer blocks and their adjacent blocks.Extensive experiments on the two publicly available datasets(i.e.,LEVIR-CD and CDD)show that the proposed LGE-Net can accurately and efficiently identify changed regions,reduce false and missed detections,and thus has a better generalization ability,compared to other state-of-the-art change detection methods.

关 键 词：深度学习 变化检测 语义分割 注意力机制 TRANSFORMER 局部特征增强 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]