当前位置:
X-MOL 学术
›
Comput. Aided Civ. Infrastruct. Eng.
›
论文详情
Our official English website, www.x-mol.net, welcomes your
feedback! (Note: you will need to create a separate account there.)
Multi‐stage detection of warped ceiling panel using ensemble vision models for automated localization and quantification
Computer-Aided Civil and Infrastructure Engineering ( IF 8.5 ) Pub Date : 2025-01-04 , DOI: 10.1111/mice.13414
Qinghua Guo, Weihang Gao, Qingzhao Kong, Xilin Lu
Computer-Aided Civil and Infrastructure Engineering ( IF 8.5 ) Pub Date : 2025-01-04 , DOI: 10.1111/mice.13414
Qinghua Guo, Weihang Gao, Qingzhao Kong, Xilin Lu
Suspended ceiling systems constitute a pivotal non‐structural component in buildings, and the warping of panels not only compromises the seismic performance but also affects the functional integrity. This paper proposes a novel multi‐stage warped panel detection (MWPD) method to automatically locate warped panels from two‐dimensional images and quantify their deformation. First, the Deep Hough Transform (DHT) is employed to localize the runner line, after that, each detected line is expanded to a rectangular strip. Then ResNet18 classifies the strips as warped or intact. Those classified as warped will undergo Gabor and horizontal Sobel filters successively to highlight the curved edge. Subsequently, the Generalized Hough Transform (GHT) is used to locate pixel points on the curve, and fitting these points yields the pixel‐level radius of curvature. Leveraging known orthogonal relationships and geometric dimensions of runners, pixel quantification is converted into physical maximum deflection. The experiments include two aspects: the first is conducted on a validation dataset to verify the localization stability, and the second is carried out on‐site for quantification validation. Results demonstrate that the proposed MWPD method effectively localizes the warped panel, achieving an accuracy of 92.2% on the validation dataset. Additionally, the quantitative test has achieved an accuracy of approximately 85%.
中文翻译:
使用集成视觉模型对翘曲的天花板进行多阶段检测,以实现自动定位和量化
吊顶系统是建筑物中关键的非结构性组成部分,面板的翘曲不仅会影响抗震性能,还会影响功能完整性。本文提出了一种新的多阶段翘曲面板检测 (MWPD) 方法,可从二维图像中自动定位翘曲面板并量化其变形。首先,采用深度霍夫变换 (DHT) 定位流道线,然后,将每条检测到的线扩展为矩形条带。然后 ResNet18 将条带分类为翘曲或完整。那些被归类为翘曲的过滤器将依次进行 Gabor 和水平 Sobel 滤光片,以突出弯曲的边缘。随后,使用广义霍夫变换 (GHT) 在曲线上定位像素点,拟合这些点可产生像素级的曲率半径。利用已知的正交关系和流道的几何尺寸,像素量化被转换为物理最大偏转。实验包括两个方面:第一个是在验证数据集上进行,以验证定位稳定性,第二个是在现场进行量化验证。结果表明,所提出的 MWPD 方法有效地定位了翘曲的面板,在验证数据集上实现了 92.2% 的准确率。此外,定量测试的准确率约为 85%。
更新日期:2025-01-04
中文翻译:

使用集成视觉模型对翘曲的天花板进行多阶段检测,以实现自动定位和量化
吊顶系统是建筑物中关键的非结构性组成部分,面板的翘曲不仅会影响抗震性能,还会影响功能完整性。本文提出了一种新的多阶段翘曲面板检测 (MWPD) 方法,可从二维图像中自动定位翘曲面板并量化其变形。首先,采用深度霍夫变换 (DHT) 定位流道线,然后,将每条检测到的线扩展为矩形条带。然后 ResNet18 将条带分类为翘曲或完整。那些被归类为翘曲的过滤器将依次进行 Gabor 和水平 Sobel 滤光片,以突出弯曲的边缘。随后,使用广义霍夫变换 (GHT) 在曲线上定位像素点,拟合这些点可产生像素级的曲率半径。利用已知的正交关系和流道的几何尺寸,像素量化被转换为物理最大偏转。实验包括两个方面:第一个是在验证数据集上进行,以验证定位稳定性,第二个是在现场进行量化验证。结果表明,所提出的 MWPD 方法有效地定位了翘曲的面板,在验证数据集上实现了 92.2% 的准确率。此外,定量测试的准确率约为 85%。