基于多尺度残差网络的隔震构造质量检测研究  

Research on quality detection of isolation detail constructions based on multiscale residual network

作　　者：党育[1] 何亚 DANG Yu;HE Ya(School of Civil Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学土木工程学院,兰州730050

出　　处：《东南大学学报(自然科学版)》2025年第1期183-193,共11页Journal of Southeast University:Natural Science Edition

基　　金：国家自然科学基金资助项目(51668043,62166025);甘肃省重点研发计划资助项目(21YF5GA073)。

摘　　要：为实现隔震构造质量的自动化检测,提出了一种基于计算机视觉的隔震构造质量检测方法。按照隔震构造图像特征和缺陷情况,将隔震构造分为7类。通过收集和拍摄全国已建的315栋隔震工程图片,构建了隔震构造数据集。参考多尺度残差网络模型Res2Net50,设计搭建了一个隔震构造质量初步检测模型ISDNet V2,该模型在Res2Net50的基础上,采用多个小卷积核堆叠,测试集结果表明:模型对各类隔震构造的识别平均准确率达到95.98%,F1分值均大于0.93,说明该模型对复杂背景的各类别隔震构造实拍图片具有很高的检测精度,检测结果偏于工程安全。对设置水平隔震缝的隔震构造,模型不仅能区别是否有缺陷,还可确定出缺陷位置。In order to achieve automatic quality detection of isolation detail constructions,this paper proposes a deep learning based isolation detail constructions quality detection method.Isolation detail constructions are classified into seven categories based on their image features and defects.An isolation detail constructions da-taset is constructed by collecting and photographing images of 315 completed isolated buildings in China.Ref-erencing the multiscale residual network model Res2Net50,an initial quality detection model for isolation de-tail constructions,ISDNet V2,is designed.This model builds upon Res2Net50 by employing multiple stacked small convolutional kernels.The results of the test set show that the average accuracy of the model in identify-ing various types of isolation detail constructions reaches 95.98%,and the F1-score is greater than 0.93,which indicates that the model has a very high accuracy in detecting pictures of various types of isolation de-tails constructions in a complex background,and the detection results are biased towards engineering safety.Additionally,for isolation detail construction with horizontal isolation seam,the model can not only distin-guish whether there are defects but also identify the location of defects.

关 键 词：多尺度残差网络 隔震构造 数据集 质量检测 

分 类 号：TU352.12[建筑科学—结构工程]