基于DeepLabV3+的建筑物高精度批量自动提取方法  

High-precision Batch Automatic Extraction of Buildings Based on DeepLabV3+

作　　者：陈梦华 张彤蕴 周子翔 陆敏 CHEN Menghua;ZHANG Tongyun;ZHOU Zixiang;LU Min(institute of Geological and Geographie information of Hunan Proince,Changsha 410000,China;Hengyang Natura!Resourees and Planning Bureau,Hengyang 420000,China)

机构地区：[1]湖南省地质地理信息所,湖南长沙410000 [2]衡阳市自然资源和规划局,湖南衡阳420000

出　　处：《测绘科学技术学报》2024年第4期375-380,共6页Journal of Geomatics Science and Technology

基　　金：湖南省地质院科研资助项目(HNGSTP202213)。

摘　　要：针对实景三维模型中建筑物提取效率低、精度不够、识别效果不理想等问题,提出一种新的基于倾斜摄影实景三维模型的建筑物批量自动提取方法。基于经典的语义分割框架DeepLabV3+训练出能从海量多样化数据中自动提取建筑物特征的DEEPROOF模型,并结合阈值聚类、降噪处理和计算机视觉中的相关图像处理方法完善提取结果。实验结果表明,与其他4种语义分割方法相比,本文方法在准确率、召回率、平均交并比上都更具优势,高出其他方法2~35个百分点;其次,分割区域更合理科学,轮廓线更清晰准确,建筑物批量自动提取的效果更好,可满足实际生产的精度要求。Due to the problems of low efficiency,insufficient precision,and unsatisfactory recognition effect of building extraction in real-scene 3D models,a new batch automatic extraction method of buildings is proposed based on 3D model of oblique photography in this paper.On the basis of the classic semantic segmentation framework DeepLabV3+,a DEEPROOF model is trained,which can automatically extract building features from massive and diverse data.Further,the extraction results are improved by combining threshold clustering,noise reduction processing,and relevant image processing methods in computer vision.The experimental results show that,compared with the other four semantic segmentation methods,the proposed method performs better in precision,recall,and mean IoU,which are 2~35 percentage points higher than those of the other methods.Moreover,the segmentation area is more reasonable and scientific,the contour lines are clearer and more accurate,and the effectiveness of automatic batch-building extractions is better,which can meet the accuracy requirements of production.

关 键 词：语义分割 实景三维模型 建筑物提取 DeepLabV3+模型 单体模型 倾斜摄影 

分 类 号：P208[天文地球—地图制图学与地理信息工程]