南方建筑Issue(4)：4-13,10.DOI:10.3969/j.issn.1000-0232.2025.04.002
街景图像深度学习驱动下的历史建筑普查与管控研究
Research on Historic Building Census and Management Method Driven by Deep Learning of Street View Images:A Case Study based on Quanzhou
摘要
Abstract
From the perspective of advancing urbanisation and renewal,historic buildings serve as crucial carriers for inheriting urban-rural contextual information and preserving the urban-rural appearance.However,it is difficult to carry out the census and management of historic buildings due to their large number and scattered distribution.It is crucial to explore a high-efficiency and convenient method that can adapt to the entire region.The emergence of artificial intelligence can help plug this gap.Combining street view big data with wide coverage and low acquisition costs,it can identify and extract traditional architectural information from images to assist surveyors in quickly screening potential historic buildings when preparing field surveys and determining the distribution characteristics of traditional building resource points on the macroscopic spatio-temporal scale.Herein,a case study based on Quanzhou,a national historical and cultural city,was carried out.Deep learning algorithms(e.g.object detection and image classification)and the GIS spatial analysis method are combined to propose an intelligent historical building recognition model composed of"establishing a traditional feature system,detecting traditional architectural features,distinguishing and screening modern buildings,and analyzing the distribution of traditional buildings"was constructed.Firstly,a traditional architectural feature system of Quanzhou—which consists of roof form,door and window styles,and decorative elements—was built according to the principles of externality,identification,and generality.It provides a basic framework for subsequent identification.Next,traditional features were recognized from all street view images of Quanzhou by using the object detection algorithm,and modern pseudo-classic architectures were then deleted using the image classification algorithm.Only street-view images of traditional buildings were retained.In addition,the accuracy of models was evaluated using the Precision,Recall,and F1 score indicators.Finally,a potential historical building map was plotted,and their spatio-temporal distribution characteristics were disclosed using ArcGIS software.In the case study,the intelligent historical building recognition model extracted 5,250 street view images of traditional buildings from 336,948 original datasets and explored 2,180 traditional building points,which were concentrated in the centre and scattered in surrounding areas.A total of 6 clusters of traditional building points were formed.In temporal evolution,the number of traditional building points generally presents a decreasing trend.According to precision evaluation and artificial verification results,the intelligent historical building recognition model has good accuracy,and its recognition efficiency is 24 times that of model data annotation and total time for model training.This study demonstrates that the intelligent historical building recognition model could extract traditional architectural features accurately and distinguish and screen modern buildings from traditional ones.The potential historical building map,which was plotted based on recognition results,exhibited quantity changes of resource points in different years and different areas.Generally speaking,the intelligent historical building recognition model has the advantages of low cost,high efficiency,stable results,and clear representation.It can develop technological utility through optimal resource allocation,fast information acquisition,and dynamic test management in the census,filing,and control of historical buildings.It adapts to high-efficiency census and dynamic monitoring demands of historical buildings in the entire region,and it provides technological support for historical building census and control practices as well as process optimization.In the future,the application of an intelligent historical building recognition model to different areas,data,and fields should be promoted to further explore its practice potential.关键词
深度学习/街景图像/历史建筑/历史保护/遗产管理/泉州Key words
deep learning/street view image/historical buildings/historical preservation/heritage management/Quanzhou分类
建筑与水利引用本文复制引用
潘莹,黄龙英,施瑛,游永熠..街景图像深度学习驱动下的历史建筑普查与管控研究[J].南方建筑,2025,(4):4-13,10.基金项目
国家自然科学基金重点资助项目(51978275):基于文化地理学的岭南汉民系传统聚落景观的特征、区划与机制研究 （51978275）
亚热带建筑与城市科学全国重点实验室自主研究课题(2023ZB09):三生视角下华南中小型海岛人居景观解析与可持续分类营建策略. （2023ZB09）
