基于对抗学习与形状修正的遥感图像建筑提取

Remote Sensing Image Building Extraction Based on Adversarial Learning and Shape Correction

摘要:

高分辨率遥感图像建筑物提取对城市实景三维重建任务具有重要意义。针对传统卷积方法分割复杂背景遥感图像中的建筑物时出现的错分、漏分和边缘混淆等问题,提出结合对抗学习与形状修正的遥感图像建筑物提取算法SCGAN。该算法引入对抗学习策略,在生成器的分割网络之后增加形状修正单元,分别通过建筑物边缘提取和形状正则化路径来提升模型对建筑物边界和形状的感知能力,并使用排除背景冗余信息、专注于建筑物形状建模的判别器来进一步指导分割网络的训练。实验结果表明,所提方法对解决遥感图像中建筑物边缘处地物互相遮挡和边界易混淆等问题具有可行性和有效性,从而整体提高了遥感图像中建筑物分割精度。

Abstract:

The extraction of buildings from high-resolution remote sensing images is of great significance for the task of three-dimensional reconstruction of urban scenes. A remote sensing image building extraction algorithm (SCGAN) is proposed, which combines adversarial learning and shape correction, to address the problems of misclassification, missed segmentation, and edge confusion that occur when traditional convolutional methods segment buildings in complex background remote sensing images. This algorithm introduces an adversarial learning strategy by adding shape correction units after the segmentation network of the generator. The model"s perception of building boundaries and shapes is improved through building edge extraction and shape regularization paths, respectively. A discriminator that excludes background redundant information and focuses on building shape modeling is used to further guide the training of the segmentation network. The experimental results show that the proposed method is feasible and effective in solving the problems of mutual occlusion and boundary confusion of ground objects at the edges of buildings in remote sensing images, thereby improving the overall accuracy of building segmentation in remote sensing images.