基于DDPM的遥感建筑轮廓数据增强方法  

Remote sensing building contour data augmentation method based on DDPM

作　　者：马孝冬 朱灵杰 解则晓[1] 高翔 MA Xiaodong;ZHU Lingjie;XIE Zexiao;GAO Xiang(College of Engineering,Ocean University of China,Qingdao 266100,China;Cenozoic Robotics,Hangzhou 310052,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国海洋大学工程学院,山东青岛266100 [2]杭州新生纪智能科技有限公司,浙江杭州310052 [3]中国科学院自动化研究所,北京100190

出　　处：《现代电子技术》2024年第21期106-112,共7页Modern Electronics Technique

基　　金：国家自然科学基金面上项目(62373349);国家自然科学基金面上项目(42076192)。

摘　　要：针对现有真实场景遥感建筑轮廓数据集存在数据丰富度有限、复杂形状数据样本量少,影响模型性能等问题,文中提出一种基于扩散模型的有限遥感数据增强方法,对合成数据和真实数据进行训练,从而得到形状更加丰富的建筑物数据,扩充建筑物数据的多样性。首先,使用扩散模型DDPM对合成建筑数据和真实场景数据训练集进行训练,生成大量与真实数据分布更接近、形状更多样的数据;然后,使用基于Transformer改进的轮廓提取模型在合成数据集上进行预训练;最后,在真实数据集上进行建筑物轮廓提取。实验结果表明:使用预训练模型训练与未使用预训练模型相比,在交并比、顶点检测精确率、顶点检测召回率、顶点检测F1分数、角度预测精确率、角度预测召回率、角度预测F1分数上分别提升了1.7%、2.4%、2.5%、2.5%、7.3%、8.2%、7.7%,表明在大规模合成数据上预训练有助于提升建筑物轮廓提取模型在真实数据上的表现;同时使用1.2×10^(5)合成数据预训练比2.4×10^(4)合成数据预训练在上述指标上提升0.8%、0.9%、1.3%、1.1%、1.1%、0.7%、0.9%,验证了预训练数据量增加对模型性能提升的有效性。In view of the limited data richness in the existing real scene remote sensing building contour data sets and the insufficient samples of complex-shaped data,which affects model performance,a limited remote sensing data enhancement method based on diffusion model is proposed.In the method,the synthetic data and real data are trained to obtain buildings with richer shapes and expand the diversity of building data.First,the denoising diffusion probabilistic model(DDPM)is used to train the synthetic building data and real scene data training set,which in turn generates a large amount of data with closer distribution and more diverse shapes than the real data.Then,the improved contour extraction model based on Transformer is used for pre-training on the synthetic dataset.Finally,the building contour extraction is performed on the real dataset.The experimental results indicate that the training with pre-trained models improves the performance in comparison with the training without pre-training on the synthetic dataset,with respective increases of 1.7%,2.4%,2.5%,2.5%,7.3%,8.2% and 7.7% in intersection over union(IoU),vertex detection precision,vertex detection recall,vertex detection F1 score,angle prediction precision,angle prediction recall,and angle prediction F1 score,which demonstrates that the pre-training on large-scale synthetic data helps improve the performance of building contour extraction models on real data.Additionally,the pre-training with 1.2×10^(5) synthetic data improves the above mentioned indicators by 0.8%,0.9%,1.3%,1.1%,1.1%,0.7% and 0.9%,respectively,in comparison with the results of the pre-training with 2.4×10^(4) synthetic data,which validates that the increase of pre-training data is effective in improving the performance of the model.

关 键 词：遥感数据 合成数据 建筑物轮廓提取 数据生成 数据增强 扩散模型 

分 类 号：TN911-34[电子电信—通信与信息系统] TP751.1[电子电信—信息与通信工程]