基于LS-CDNet的轻量级遥感影像变化检测  

Lightweight remote sensing image change detection based on LS-CDNet

作　　者：黄子岸 赵艮平 王卓薇[1] HUANG Zian;ZHAO Genping;WANG Zhuowei(Guangdong Provincial Key Laboratory of Cyber Physical System,School of Computer Science,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广州510006

出　　处：《激光杂志》2024年第10期67-73,共7页Laser Journal

基　　金：国家自然科学基金资助项目(No.61701123);广州市基础与应用研究项目(No.202201010273);教育部与北京航天宏图信息技术有限公司协同教育项目(No.220802313160749);广东省农业科学与技术创新团队项目(No.2019KJ147)。

摘　　要：为实现高效的遥感影像变化检测,提出一种基于轻量级孪生网络的变化检测方法(A Lightweight Siamese Network for Change Detection,LS-CDNet)。该方法以轻量级网络MobileNetV2为特征提取网络,基于孪生网络架构进行构建。设计一个串联注意力模块对低层特征进行优化以增强区域边界特征信息。综合考虑模型训练的收敛效率及对正负样本不均衡的数据集学习能力,引入损失函数BCE Loss和Dice Loss加权组合以改进模型学习策略。在LEVIR-CD和CDD两个数据集的实验结果表明,LS-CDNet精确率分别达到88.06%和90.12%,模型参数量为3.76 M,计算量为2.18 G(FLOPs),性能表现优于其他对比方法。To achieve efficient change detection in remote sensing images,a lightweight Siamese network for change detection(LS-CDNet)was proposed.LS-CDNet is constructed based on a Siamese network architecture,with the lightweight network MobileNetV2 used as the backbone.A cascaded attention module is designed to optimize the low-level features and enhance the boundary information of the regions.In order to consider both the convergence efficiency of model training and the learning ability of imbalanced datasets for positive and negative samples,a weighted combination of the BCE Loss and Dice Loss is used to optimize the model learning strategy.Experimental results on the LEVIR-CD and CDD datasets demonstrate that LS-CDNet achieves precision rates of 88.06%and 90.12%respectively,with a model parameter size of 3.76 M and a computational cost of 2.18 G(FLOPs).The performance of LS-CDNet outperforms other comparative methods.

关 键 词：变化检测 轻量化 注意力机制 孪生网络 遥感影像 

分 类 号：TN249[电子电信—物理电子学]