摘要
针对古建筑脊兽识别准确率和自动化程度不高的问题,本文提出一种基于卷积神经网络的脊兽自动识别方法。该方法主要分为4个步聚:①爬取众源数据建立脊兽数据集;②构建脊兽特征金字塔网络(Ridge Beast-Feature Pyramid Network,RB-FPN)提取图像深度语义,检测脊兽的潜在区域;③利用ImageNet预训练权重精调ResNet50模型参数,实现脊兽种类精细分类;④识别测试集样本综合评价性能指标。试验结果表明,本文所提出的脊兽自动识别方法准确率可达92.17%,召回率为82.02%,F 1值为0.87,能有效地识别数字图像中的脊兽装饰件种类,结果可满足屋顶精细三维重建、维护管理与断代等应用需求。
To improve the accuracy and automation degree of ridge beast recognition,we proposed an approach to recognize the ridge beast based on convolutional neural network.Four steps are involved.Firstly,we built the ridge beast dataset by data crawling.Then we derived the deep semantic information by ridge beast-feature pyramid network.The parameters of ResNet50 was adjusted by ImageNet pretraining weights,and the fine-grained classification of ridge beasts was achieved.Finally,we tested the feasibility and suitability of training sets.Results indicated that the accuracy of ridge beasts derived by the proposed algorithm reaches 92.17%,with recall ratio as 82.02%and F1 value as 0.87.The algorithm can be used for fined 3D reconstruction maintenance and backtracking of roofs.
作者
纪宇航
董友强
侯妙乐
齐莹
霍芃芃
JI Yuhang;DONG Youqiang;HOU Miaole;QI Ying;HUO Pengpeng(School of Geomatics and Urban Spatial Informatics,Beijing University of Civil Engineering and Architecture,Beijing 102616 China;Beijing Key Laboratory for Architectural Heritage Fine Reconstruction&Health Monitoring,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;School of Architecture and Urban Planning,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Beijing Institute of Surveying and Mapping,Beijing 100089,China)
出处
《地理信息世界》
2021年第3期54-60,共7页
Geomatics World
基金
国家重点研发计划项目(2019YFC1520800)
北京市属高校高水平教师队伍建设支持计划长城学者培养计划项目(CIT&TCD20180322)
北京市教委科研计划项目(KM202110016005)
北京市自然科学基金项目——市教委联合基金项目(KZ202110016021)
北京建筑大学研究生创新项目(PG2020079)。
关键词
卷积神经网络
目标检测
图像分类
脊兽
convolutional neural network
objective detection
image classification
ridge beast