基于深度学习的灌区水工混凝土结构表面裂缝自动检测方法  

Automatic detection method of surface cracks of hydraulic concrete structures in irrigation area based on deep learning

作　　者：曹国金 苏超[2] 王文君 Cao Guojin;Su Chao;Wang Wenjun(Guangzhou Liuxihe Irrigation District Management Center,Guangzhou510920,Guangdong;School of Water Conservancy and Hydropower,Hehai University,Nanjing210024,Jiangsu)

机构地区：[1]广州市流溪河灌区管理中心,广东广州510920 [2]河海大学水利水电学院,江苏南京210024

出　　处：《吉林水利》2022年第12期14-17,共4页Jilin Water Resources

基　　金：2021年度广州市水务科技项目:广州市流溪河灌区干渠设施智能化管理关键技术研究。

摘　　要：裂缝是水利灌区混凝土结构破坏损伤的主要表现形式之一,裂缝的存在破坏了结构的整体性,削弱了结构承载力。传统的现场检查主要依靠人工视觉,主观程度高。本文提出了一种基于深度学习的混凝土结构表面裂缝自动检测方法,利用手机照片或无人机巡查照片通过图片识别技术确定混凝土表面裂缝位置、并在图片上做出标记。方法构建了一个单阶段目标检测网络模型,它由主干特征网络EfficientNet和检测器两部分组成。检测器从三个不同尺度的特征图中提取特征要素作为输入,并将低层特征与高层特征融合,提高了检测精度。将该模型与其他三种目标检测模型进行了比较,本文模型平均检测精度最高,能够很好地检测出图像中的裂缝。Crack is one of the main forms of concrete structure damage in water conservancy irrigation area,the existence of crack destroys the integrity of the structure and weakens the bearing capacity of the structure.The traditional on-site inspection mainly depends on artificial vision with a high degree of subjectivity.This paper proposes an automatic detection method of concrete structure surface cracks based on deep learning,using mobile phone photos or UAV patrol photos,through image recognition technology to determine the location of concrete surface cracks,and make marks on the pictures.This method constructs a single-stage target detection network model,which consists of two parts:the backbone feature network EfficientNet and the detector.The detector extracts feature features from three feature images of different scales as input,and fuses low-level features with high-level features to improve the detection accuracy.After comparing this model with the other three target detection models,it is found that the average detection accuracy of this model is the highest and can better detect cracks in the image.

关 键 词：灌区 水工混凝土 裂缝检测 深度学习 目标检测 

分 类 号：TV544[水利工程—水利水电工程] TV698.1