摘要
针对遥感图像人工标注耗时且昂贵的缺点,提出一种两阶段的变化检测方法。通过预训练去噪扩散概率模型来利用这些现成的、未标注的遥感图像信息,利用从扩散模型主干网络U-Net的后半部分编码器中获取的多尺度特征来训练一个轻量级的变化检测头部。通过同时处理不同加噪时间步的遥感图像,基于噪声水平进行加权融合进一步提升模型对变化相关信息的敏感性。在LEVIR-CD和WHU-CD数据集上的对比实验结果表明,该方法有效提高了识别精度。
A two-stage change detection method was proposed to address the drawbacks of manual labeling of remote sensing images,which is time-consuming and expensive.This readily available,unlabeled remote sensing image information was utilized by pre-training a denoised diffusion probabilistic model to train a lightweight change detection head using multi-scale features obtained from the encoder of the second half of the diffusion model backbone network,U-Net.The sensitivity of the model to change-related information was further enhanced by simultaneously processing remote sensing images with different noise addition time steps and weighted fusion based on the noise level.Comparative experiments on LEVIR-CD and WHU-CD datasets show that the method effectively improves the recognition accuracy.
作者
李克文
蒋衡杰
李国庆
姚贤哲
刘文龙
LI Ke-wen;JIANG Heng-jie;LI Guo-qing;YAO Xian-zhe;LIU Wen-long(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580,China)
出处
《计算机工程与设计》
北大核心
2025年第2期337-344,共8页
Computer Engineering and Design
基金
国家自然科学基金重大基金项目(51991365)
山东省自然科学基金项目(ZR2021MF082)。
关键词
变化检测
深度学习
预训练
特征融合
特征提取
扩散模型
无监督训练
change detection
deep learning
pre-training
feature fusion
feature extraction
diffusion modeling
unsupervised training
