| 摘要: |
| 近年来，基于卷积神经网络（Convolutional Neural Network,CNN）的方法尤其是孪生网络成为遥感图像变化检测任务的主流网络。然而，传统孪生网络模型在特征提取与表征能力上存在固有局限，难以适应变化区域在几何形态和空间尺度上的多样性，导致检测结果常出现伪变化误判与真实变化漏检。针对上述技术瓶颈，提出了一种高性能像素级变化检测模型MNUNet-CD（Mix-Nested-Unet Change Detection）。该模型通过混合残差模块对双时相图像对进行多维度特征提取，构建层次化特征表达体系；引入多尺度特征融合机制，实现对不同空间分辨率下变化模式的精细化捕捉；设计全局注意力模块对特征空间进行自适应筛选，强化模型对关键变化特征的表征能力。实验结果表明，所提模型在CDD、WHU-CD和LEVIR-CD 3个公开数据集上均展现出一定性能优势，与基准方法(SNUNet-CD)相比，在F1分数上分别提升了 3.3%、0.7% 和 0.9%。 |
| 关键词: 遥感图像 图像变化检测 深度学习 孪生网络 全局注意力 |
| DOI：10.20079/j.issn.1001-893x.250428001 |
|
| 基金项目: |
|
| Change Detection of Remote Sensing Image Based on Mix Residual and Global Attention |
| LI Zhao,XU Tao,TIAN Xilan |
| (The 38th Research Institute of China Electronics Technology Group Corporation，Hefei 230088，China) |
| Abstract: |
| In recent years,methods based on convolutional neural network(CNN),especially the siamese network,have become the mainstream networks for remote sensing image change detection tasks.However,traditional siamese network models have inherent limitations in feature extraction and representation capabilities,making it difficult to adapt to the diversity of geometric shapes and spatial scales in change regions,leading to common issues such as false change misjudgments and missing real changes in detection results.To address the above technical bottlenecks,a high-performance pixel-level change detection model,Mix-Nested-Unet Change Detection(MNUNet-CD),is proposed.The model employs a Mix Residual Block for multi-dimensional feature extraction on dual-temporal image pairs,constructing a hierarchical feature representation system.It introduces a multi-scale feature fusion mechanism to achieve fine-grained capture of change patterns at different spatial resolutions.Additionally,a global attention module is designed to adaptively filter the feature space,enhancing the model’s ability to represent key change features.Experimental results show that the proposed model demonstrates certain performance advantages on three public datasets(CDD,WHU-CD,and LEVIR-CD).Compared with the benchmark method(SNUNet-CD),it achieves F1 score improvements of 3.3%,0.7%,and 0.9%,respectively. |
| Key words: remote sensing image image change detection deep learning siamese network global attention |
