基于区域候选网络的矢量线要素自动化简方法  

Automatic vector polyline simplification based on region proposal network

作　　者：江宝得 许少芬 巫勇 王淼 JIANG Baode;XU Shaofen;WU Yong;WANG Miao(School of Computer Science,China University of Geosciences,Wuhan 430074,China;National Engineering Research Center of Geographic Information System,China University of Geosciences,Wuhan 430074,China)

机构地区：[1]中国地质大学(武汉)计算机学院,湖北武汉430074 [2]中国地质大学(武汉)国家地理信息系统工程技术研究中心,湖北武汉430074

出　　处：《测绘学报》2023年第12期2209-2222,共14页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42171408)。

摘　　要：针对现有矢量线要素化简算法智能化程度不高的问题,本文提出了一种基于区域候选网络的矢量线要素自动化简方法。首先利用深度可分离卷积神经网络实现矢量线要素栅格化后的卷积特征提取;然后结合线要素的坐标信息改进区域候选网络中候选框的生成方式,以实现候选框与可能的弯曲组合的对应;最后在统一线要素弯曲特征图大小的基础上,根据候选框对应的卷积特征进行二分类判断,完成弯曲单元的自动识别,并通过删除弯曲单元实现线化简。本文在海岸线数据集上进行了模型的训练和测试,并通过不同栅格化参数、不同骨干网络、跨尺度化简,以及不同类型线要素化简等对比试验验证了模型的有效性。试验结果表明,本文方法能够自动学习已有线化简案例中的化简知识,并充分利用线要素的矢量特征及栅格特征,自动完成线要素弯曲特征的准确定位,最终实现可端到端训练的矢量线要素自动化简。To address the problem of existing vector polyline simplification algorithms lacking intelligence,an automatic vector polyline simplification algorithm based on regional proposal network is proposed.Firstly,a depth-separable convolutional neural network is used to realize the convolutional feature extraction of raster polylines.Then,the generation method of region proposals in the region proposal network is improved by combining the coordinate information of the polylines,which is in order to establish the correspondence between the proposal region and the possible bending combinations.Finally,after unifying the size of the bending feature map,the automatic detection of bending units is completed by binary classification according to the convolutional features corresponding to the region proposals,and polyline simplification is achieved by deleting the bending units.In this paper,the model is trained and tested on the coastline dataset,and the effectiveness of the model is verified by comparison experiments with different rasterization parameters,different backbone networks,cross-scale simplification,and different types of line simplification.Experimental results show that this algorithm can intelligently learn the simplification knowledge from the existing polyline simplification cases,and make full use of the vector and raster features of polylines to identify and locate the bending units,and finally complete the automatic simplification of vector polylines while the network can be trained end-to-end.

关 键 词：线化简 区域候选网络 地图综合 深度学习 

分 类 号：P283[天文地球—地图制图学与地理信息工程]