边缘监督与特征融合的遥感影像建筑物提取方法

Studying on building extraction method for remote sensing images combined with multi-scale edge supervision and high and low feature attention fusion

针对现有深度学习模型的建筑物提取结果普遍存在的边缘规整度差、细节不丰富等问题,该文提出了一种结合边缘监督与特征融合的深度学习建筑物提取模型,此模型以U-Net为基本架构,借助多任务学习思想,设计了多尺度边缘监督分支和高低层特征注意力融合模块,有效提升了建筑物提取结果的边缘规整程度和细节丰富程度。在WHU建筑物数据集和Massachusetts建筑物数据集上进行对比实验,并进行了建筑物提取和建筑物边缘的精度检验。结果表明,与SegNet、U-Net、Attention U-Net、D-LinkNet、STTNet模型相比,该文的方法能有效地提升各指标精度,提取出的建筑物边缘更加准确、细节更加丰富,并且有效地提升了大型建筑物提取的完整性。

Regarding the common issues of poor edge regularity and lack of detailed information in the building extraction results produced by existing deep learning models.This paper proposes a deep learning-based building extraction model that combines edge supervision and feature fusion.The model leverages U-Net as its fundamental architecture.It designs a multi-scale edge supervision branch and a high and low feature attention fusion module,effectively improving the edge regularity and level of detail in the building extraction results.The experiments in the WHU Building Dataset and the Massachusetts Building Dataset are conducted to compare the performance with other models including SegNet,U-Net,Attention U-Net,D-LinkNet,and STTNet.Accuracy tests for building extraction and building edge extraction are also performed.The results show that compared to the mentioned models,our proposed method significantly improves the accuracy of various metrics,generates more accurate and detailed building edges,and enhances the completeness of large building extraction.