This paper presents findings of delamination detection using infrared thermography (IRT) in five in-service bridges using an unmanned aerial vehicle system. The authors have used semantically segmented IRT images to evaluate IRT's effectiveness in detection of deck delamination for the first time. Using an adaptive image processing-based model, sub-surface delaminations were detected by optimizing all user-defined parameters in the model, including threshold values to convert the enhanced IRT images to a binary image. The optimization process has been done selecting iterating the user-defined parameters and their effect on the interaction of a set of sigmoid curves representing the model's performance metrics. The 2-clustered (Park river median and Park river south bound bridges) and 3-clustered (Park river north-bound, Forest river north-bound, Forest river south-bound) threshold values ranged from 0.365 to 0.380 and 0.459 to 0.486, respectively, and yielded to an average accuracy of 69% for delamination detection. The effect of different parameters on the value of the performance metrics were investigated and analyzed including the ambient wind speed and depth of delamination during data collection. The optimized delamination detection model was shown to be superior to a delamination detection using the conventional unsupervised K-nearest neighbor clustering technique.

本文介绍了使用无人机系统在五个在役桥梁中使用红外热成像 (IRT) 进行分层检测的结果。作者首次使用语义分割的 IRT 图像来评估 IRT 在检测甲板分层方面的有效性。使用基于自适应图像处理的模型，通过优化模型中所有用户定义的参数（包括将增强的 IRT 图像转换为二值图像的阈值）来检测亚表面分层。优化过程已经完成，选择迭代用户定义的参数及其对代表模型性能指标的一组 sigmoid 曲线的交互作用的影响。2 集群（帕克河中线和帕克河南行桥）和 3 集群（帕克河北行，森林河流北行、森林河流南行）阈值范围分别为 0.365 至 0.380 和 0.459 至 0.486，分层检测的平均准确率为 69%。研究和分析了不同参数对性能指标值的影响，包括数据收集过程中的环境风速和分层深度。优化的分层检测模型被证明优于使用传统无监督 K-最近邻聚类技术的分层检测。研究和分析了不同参数对性能指标值的影响，包括数据收集过程中的环境风速和分层深度。优化的分层检测模型被证明优于使用传统无监督 K-最近邻聚类技术的分层检测。研究和分析了不同参数对性能指标值的影响，包括数据收集过程中的环境风速和分层深度。优化的分层检测模型被证明优于使用传统无监督 K-最近邻聚类技术的分层检测。