基于多边特征引导聚合网络的变化检测算法  

Change Detection Algorithm Based on Multilateral Feature Guided Aggregation Network

作　　者：冯星宇 朱灵龙 张永宏[1,2,3] 阚希 曹海啸 马光义 FENG Xingyu;ZHU Linglong;ZHANG Yonghong;KAN Xi;CAO Haixiao;MA Guangyi(School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Internet Engineering,Wuxi University,Wuxi,Jiangsu 214105,China;Collaborative Innovation Center for Atmospheric Environment and Equipment Technology,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学自动化学院,南京210044 [2]无锡学院物联网工程学院,江苏无锡214105 [3]南京信息工程大学江苏省大气环境与装备技术协同创新中心,南京210044

出　　处：《计算机工程与应用》2025年第3期264-274,共11页Computer Engineering and Applications

基　　金：国家自然科学基金(42305158,42105143);江苏省高等学校基础科学(自然科学)研究面上项目(23KJB170025);无锡市“太湖之光”基础研究项目(K20231021);无锡学院人才启动项目(2022r035)。

摘　　要：现有的遥感图像变化检测方法主要依赖卷积神经网络(convolutional neural network,CNN)或Transformer进行构建,但这些方法通常未能充分平衡这两种技术的优缺点,并且往往没有专门针对变化检测的任务特性(对变化区域特征信息进行提取学习)进行优化设计。针对这一问题,充分利用了Transformer的全局信息处理能力和CNN的局部信息捕获能力,提出了一种充分结合两者各自优势并由多条支路组成的多边特征引导聚合网络模型,该模型通过基于Transformer的主网络来对图像的全局信息进行提取,通过设计的基于CNN的多尺度特征提取模块来对图像的局部信息进行提取,通过特征聚合网络将图像的全局信息分别与变化和未变化区域信息进行聚合后输出得到预测结果图。为了验证模型的有效性,构建了一个包含多种地表覆盖类型,涵盖不同季节的新的遥感图像变化检测数据集。同时在实验部分也利用两个公开数据集来进一步验证了模型的泛化性和鲁棒性。实验结果表明,与现有先进方法相比,所提算法在三个数据集上的平均交并比(mean intersection over union,MIoU)指标分别提高了0.83、0.71、0.7个百分点。Existing remote sensing image change detection methods primarily rely on convolutional neural network(CNN)or Transformer for development.However,these methods often fail to fully balance the advantages and disadvantages of these two technologies and typically do not optimize design specifically for the task characteristics of change detection(extracting and learning features of change regions).Addressing this issue,this paper fully utilizes the global information processing capability of Transformer and the local information capture ability of CNN,proposing a multilateral feature guidance aggregation network model that combines the strengths of both.The model first extracts global information from images through a Transformer-based main network,then captures local information through a designed multi-scale CNN-based feature extraction module,and finally aggregates the global information with the changed and unchanged area information through a feature aggregation network to output predicted result images.Additionally,to verify the effectiveness of model,a new remote sensing image change detection dataset is constructed,which includes various land cover types and changes across different seasons,aiming to provide a more comprehensive testing scenario.The experimental section also uses two public datasets to further verify the generalizability and robustness of model.Experimental results show that compared to existing advanced methods,the proposed algorithm improves the mean intersection over union(MIoU)metrics by 0.83,0.71,and 0.7 percentage points on the three datasets,respectively.

关 键 词：遥感图像 变化检测 卷积神经网络 TRANSFORMER 特征聚合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] X705[自动化与计算机技术—计算机科学与技术]