PDF全文阅读

冯志慧¹，曹文凯¹，薛鹏涛²，梁志强³

 

 

(1.河南农业大学 信息与管理科学学院，郑州 450002；2.河南省交通事业发展中心，郑州 450016； 3.长安大学 公路学院，西安 710064)

 

 

摘要：为改进桥梁病害识别效率和质量，提出一种基于贝叶斯优化和卷积神经网络的深度学习算法，进行桥梁麻面、裂缝、露筋和剥落等4种桥梁病害识别；针对裂缝病害，建立以Mobilenet-v2特征提取网络的DeepLabv3+作为裂缝图像语义分割模型。结果表明:基于贝叶斯优化和卷积神经网络的深度学习算法对4种桥梁病害识别精度及鲁棒性基本保持不变，训练时间减少了约80%；通过裂缝图像语义分割模型和图像处理技术实现裂缝的精准分割和几何信息的自动提取计算，分割裂缝MIoU达到0.95。桥梁病害的高效和精准识别，为桥梁性能预测分析提供了更加准确的数据参考。

关键词：深度学习；病害识别；裂缝特征；卷积神经网络；贝叶斯优化；语义分割

中图分类号：TP391

文献标志码：A

文章编号：1005-8249（2025）06-0140-10

DOI：10.19860/j.cnki.issn1005-8249.2025-06-024

 

 

FENG Zhihui¹,CAO Wenkai¹,XUE Pengtao²,LIANG Zhiqiang³

 

 

(1.Henan Agricultural University College of Information and Management Science,Zhengzhou 450002,China；

 

2.Transportation Development Center of Henan Province,Zhengzhou 450016,China；

 

3.Chang’an University Highway College,Xi’an 710064,China)

 

Abstract：To improve the efficiency and accuracy of bridge defect identification,a deep learning algorithm based on bayesian optimization and convolutional neural networks is proposed for identification of four types of bridge defects:pockmarked surface,cracks,exposed rebar,and spalling.For cracks,DeepLabv3+ with Mobilenet-v2 as the feature extraction network is established as the semantic segmentation model for crack images.The results show that the proposed method maintains nearly the same accuracy and robustness in identifying the four types of bridge defects,while reducing training time by approximately 80%.Through the semantic segmentation model and image processing techniques,precise segmentation of cracks and automatic extraction of geometric information are achieved,with an MIoU of 0.95 for crack segmentation.The efficient and accurate identification of bridge diseases provides more precise data references for bridge performance prediction and analysis.

 

Key words：deep learning;damage identification;crack features;convolutional neural network;Bayesian optimization;semantic segmentation

基金项目：国家自然科学基金项目(51878059)，河南省科技攻关项目(252102240037)。

作者简介：冯志慧(1977—)，女，博士，副教授。研究方向:机器学习。

通信作者：薛鹏涛(1978—)，男，博士。研究方向:桥梁工程。

收稿日期：2023-04-07