一种双重深度特征监督变化检测网络  

A dual deep feature supervised change detection network

作　　者：韩承熙 苏晓玉 魏志强 HAN Chengxi;SU Xiaoyu;WEI Zhiqiang(Intelligent Science&Technology Academy Limited of China Aerospace Science and Industry Corporation,Beijing 100043,China)

机构地区：[1]中国航天科工集团智能科技研究院有限公司,北京100043

出　　处：《航天工程大学学报》2025年第2期62-69,共8页

摘　　要：深度学习方法在遥感影像变化检测中取得了明显进展,但在边缘信息提取和建筑物内部细节提取方面仍面临挑战,现有方法在提取建筑物边缘和内部细节信息时经常出现模糊和空洞现象,为解决这一问题,提出了一种双重深度特征监督变化检测网络(Dual Deep Feature Supervised Change Detection Network,DDDNet)。通过两轮深度特征监督实现多尺度的变化特征信息融合,提高模型对建筑物边缘和内部特征的表达,引入变化特征改进的注意力机制模块,引导模型关注建筑物边缘与内部区域信息,弥补了信息提取不完整的情况。在LEVIR-CD、WHU-CD、SYSU-CD和S2 Looking-CD4个挑战性数据集上通过定量与定性的方式验证了所提模型的有效性。Deep learning methods have made significant progress in remote sensing image change detection,but they still face challenges in edge information extraction and building interior details extraction.The existing methods often suffer from blurring and hollowing phenomena when extracting the edge information and internal details.To solve these problems,this paper proposes a dual deep feature supervised change detection network(DDDNet).The multiscale change feature information fusion is realized by two rounds of deep feature monitoring,so as to improve the expression of the edge and internal features of the building.A feature improvement module with an attention mechanism module improved by changing features is introduced to guide the model to pay attention to the edge and internal area information of buildings to make up for the incomplete information extraction.The effectiveness of the proposed model is verified quantitatively and qualitatively on four challenging datasets:LEVIR-CD,WHU-CD,SYSU-CD and S2 Looking-CD.

关 键 词：遥感影像 深度学习 变化检测 深度监督 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]