首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 238 全文下载次数: 163
引用本文:

DOI:

10.11834/jrs.20219335

收稿日期:

2019-09-20

修改日期:

2019-12-29

PDF Free   EndNote   BibTeX
基于边界约束网络和分水岭分割算法的建筑提取
罗壮1, 李明2, 张德朝3
1.太原理工大学;2.太原理工大学大数据学院;3.清华大学地球系统科学系
摘要:

城市作为高密度建筑区域,在较小范围内有大量结构相似的建筑紧密分布。当前从高分辨率图像中准确检测建筑仍然是一个挑战,本文受边缘检测网络启发,提出一种强化边界精度的建筑物提取新方案,根据建筑物及边界特点改进深度网络,结合自下而上分组的分水岭分割提高分类精度和建筑边界的准确度。首先对数据预处理,生成建筑边界和建筑分割线两类辅助标签;改进性能较优的建筑检测框架ICT-Net网络,修改网络结构和损失函数,针对两类辅助标签,强化边界影响,提高网络性能;最后对网络预测结果应用结合分水岭分割和梯度提升回归树的后处理,实现高精度的建筑提取。结果表明,数据预处理、改进深度学习算法可提高建筑检测像素精度Intersection over Union(IOU)约1%。后处理能充分利用网络输出的概率信息,有效优化建筑边界,在网络预测结果的基础上提高建筑实例召回率10.5%。本文方案与原始的ICT-Net网络相比,提高建筑实例召回率22.9%。

Building detection based on boundary regulated network and watershed segmentation algorithm
Abstract:

Object High-density urban cities contain numerous similar buildings positioned in close proximity. Building detection from high spatial resolution remote sensing imagery in such scenes remains a challenge in the fields of computer vision and remote sensing urban application. The integration of traditional segmentation algorithms and a novel neural network is an effective way for such challenging settings. Inspired by the recent success of deep-learning-based edge detection, a new building detection method aiming at accurate boundary is proposed. According to characteristics of building and its border, this paper improves the network structure and integrates the network with bottom-up watershed segmentation to improve boundary precision as well as classification accuracy. Method First, two auxiliary labels-the building boundary and the parting line are derived from the original dataset through data preprocessing. Subsequently, a newly proposed building detection frame ICT-Net is improved by modifying structure and loss function in accordance with two auxiliary labels to obtain the probability of three classes. Finally, one postprocess integrating watershed segmentation with gradient boosted regression trees is employed to achieve high accuracy of building detection. Specifically, the probability feature map is generated by merging the probability of three classes. Watershed segmentation with building marker thresholds is applied to obtain building instances from probability feature map. Then the building probability of each building instance predicted by gradient boosted regression trees is used to select building instances, resulting in building detection results. In addition, parameters selection is implemented. Result The performance of the proposed method is validated on INRIA dataset which provides aerial orthorectified color imagery with a spatial resolution of 0.3m with corresponding ground truth label for two semantic classes: building and not building. The experimental results suggest that data preprocessing and the application of boundary loss can gain an improvement of 1% in terms of Intersection over Union (IOU) of building detection. Postprocess can take full advantage of probability information from the network, effectively optimizing the building boundary. Postprocess brings an improvement of 10.5% in terms of building instance recall compared with that of results from the neural network. Our paper has a building instance recall that is 22.9% higher than that of the original ICT-Net. Conclusion A novel building detection method based on boundary regulated network and watershed segmentation is proposed in this study. The experimental results show the advantages of the enhanced-boundary-oriented data preprocessing and the modified neural network and demonstrate the proposed method can further improve prediction accuracy on the basis of network. Nevertheless, the great performance of the proposed method largely depends on parameters selection, and further improvements should be made in the future.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮