摘要
地图目标的形状在地图制图综合、空间查询等研究中发挥着重要作用。地图建筑物形状的识别与分类作为建筑物轮廓化简与典型化的基础,一直是制图综合研究的热点问题。目前,主要的建筑物形状识别方法主要依赖对建筑物轮廓的描述,对建筑物等地图面状要素的形态特征有较强的依赖性,通常只在应对特定类型的规则轮廓或直角化轮廓时能发挥较好的效果,对于形状不规则或复杂的情况识别不佳。本文提出一种AlexNet支持下的地图建筑物形状分类方法,将矢量地图中建筑物数据的形状分类问题,转化为建筑物栅格图像的分类问题,通过完成卷积神经网络的图形分类实现建筑物的形状识别。该方法首先结合空间认知规律提出一系列典型建筑物形状类型,然后利用矢量-栅格转换的方法从OSM数据采样单体建筑物栅格图像,通过人工标识获得建筑物形状分类训练样本,训练AlexNet卷积神经网络分类模型,最后利用训练好的模型对大比例尺建筑物数据进行智能形状分类与识别。本文利用北京、香港2个城市的OSM建筑物数据作为样本训练建筑物形状分类模型,并在广州部分城区的OSM建筑物数据上进行验证。相较传统形状相似性度量方法,本文提出的方法对实验区建筑物的识别分类总体查全率提高了2.48%,达到92.32%,对于较为复杂的形状(如T形、十字形)识别也具有更高的精度,查准率分别提高了13.83%和24.53%。实验结果表明本文提出的方法对建筑物形状分类的效果有明显提升,能够实现常见建筑物形状的有效分类,为下一步的建筑物化简、典型化等综合操作打下了基础。
Shape of map objects plays an important role in the study of map generalization and spatial query. As the basis of simplification and typification of building, the recognition and classification of map building shapes has always been a hot issue in cartographic generalization research. At present, the traditional building shape recognition methods mainly rely on the description of the building boundary and a specific shape similarity calculation, which can only be applied to buildings with conventional shapes. The traditional methods have a strong dependence on the morphological characteristics of map surface elements such as buildings, and usually only play a good role in dealing with specific types of regular contours or rectangular contours, but has poor shape recognition ability for buildings with complex or unusual shapes. This study proposes a new method of map building shape classification method based on AlexNet. The shape classification problem of building data in vector map is transformed into the classification problem of building raster images, and the shape recognition of building is realized by completing the graphic classification of convolutional neural network. Firstly, this method constructs a series of typical shape types based on spatial cognition. Secondly, the raster images of individual buildings are sampled from OSM data by vector-raster transformation method, and the training samples of building shape classification are obtained through manual identification. Based on this, the classification model of AlexNet convolutional neural network is trained. Thirdly, this method uses the trained model to perform intelligent shape classification and recognition on large-scale building data. In this paper, the OSM building data of Beijing and Hong Kong were used as samples to train the building shape classification model, and the proposed method was verified using the OSM building data of some urban areas in Guangzhou. Compared with the traditional shape similarity measurement method, the recall rat
作者
焦洋洋
刘平芝
刘爱龙
刘松林
JIAO Yangyang;LIU Pingzhi;LIU Ailong;LIU Songlin(Information Engineering University,Zhengzhou 450001,China;State Key Laboratory of Geo-information Engineering,Xi'an,710054,China;Xi'an Research institute of Surveying and Mapping,Xi'an 710054,China)
出处
《地球信息科学学报》
CSCD
北大核心
2022年第12期2333-2341,共9页
Journal of Geo-information Science
基金
国家自然科学基金项目(42071450、41801396、62101395)。
关键词
制图综合
建筑物识别
深度学习
卷积神经网络
模板匹配
形状分类
建筑物化简
矢栅结合
cartographic generalization
building recognition
deep learning
convolutional neural network
template matching
shape classification
building simplification
vector-raster combination
