基于深度学习的混凝土结构钢筋工程质量图像视觉检测算法  

Image visual detection method for rebar engineering quality in concrete structure based on deep learning

作　　者：杜晓庆[1] 葛潇峰 朱炯亦 汪德江[1] 蒋海里 刘攀攀 DU Xiaoqing;GE Xiaofeng;ZHU Jiongyi;WANG Dejiang;JIANG Haili;LIU Panpan(School of Mechanics and Engineering Science,Shanghai University,Shanghai 200444,China;Shanghai Highway and Bridge Group Co.,Ltd.,Shanghai 200082,China)

机构地区：[1]上海大学力学与工程科学学院,上海200444 [2]上海公路桥梁(集团)有限公司,上海200082

出　　处：《建筑科学与工程学报》2024年第6期31-40,共10页Journal of Architecture and Civil Engineering

基　　金：国家自然科学基金项目(51978392);国家自然科学基金青年科学基金项目(52208317);上海市科技攻关基金项目(20dz1202100);上海市促进产业高质量发展专项项目(2023-GZL-RGZN-01031)。

摘　　要：针对目前钢筋工程质量检测中人工抽检方法检测效率低,覆盖面不全的问题,提出一种可实现钢筋间距与直径智能检测的改进YOLOX目标检测算法。该方法基于精确的钢筋交叉点识别,实现钢筋间距与直径检测,在YOLOX目标检测算法中加入坐标注意力机制模块,采用完全交并比损失函数替换原有算法中的边界框回归损失函数,显著提升目标检测算法识别钢筋交叉点及其中心坐标的精准度;依据钢筋交叉点预测框的像素坐标信息与RGBD相机采集的深度信息,实现钢筋间距检测;采用整体嵌套边缘检测网络(HED)边缘检测算法消除图片中钢筋肋边缘对统计钢筋直径所含像素个数的干扰,实现钢筋直径的检测。结果表明:采用改进后的算法检测得到的钢筋间距最大误差为4.04 mm,平均误差小于2.8 mm,满足施工规范要求;当采用改进后的算法检测8、10、16、20、25 mm钢筋直径时,检测值在标准值d0±1 mm范围内的平均准确率为97.22%。Aiming at the problems of low detection efficiency and incomplete coverage caused by the current manual sampling method for quality inspection of rebar engineering,an improved YOLOX object detection algorithm was proposed to achieve intelligent detection of rebar spacing and diameter.The method realized rebar spacing and diameter detection based on accurate rebar intersection identification.The coordinate attention mechanism module was added to YOLOX object detection algorithm,and the original loss function for bounding box regression was replaced with the complete intersection over union loss function,which significantly improved the accuracy of the object detection algorithm in detecting both the rebar intersection and its central coordinates.According to the pixel coordinate information from the rebar intersection prediction box,along with depth information collected by the RGBD camera,the rebar spacing detection was realized.The Holistically-nested Edge Detection(HED)edge detection algorithm was used to eliminate the interference of rebar rib edges in the image on the number of pixels included in the statistical rebar diameter,to realize the detection of rebar diameter.The results show that the maximum error of rebar spacing detected by the improved algorithm is 4.04 mm,and the average error is less than 2.8 mm,which meets the requirements of construction specifications.When the improved algorithm is used to detect 8,10,16,20,25 mm rebar diameters,the average accuracy of the detection value in the range of standard value d 0±1 mm is 97.22%.

关 键 词：钢筋工程质量 深度学习 目标检测算法 钢筋交叉点 像素坐标 

分 类 号：TU317[建筑科学—结构工程]