Reinforced concrete bridge substructures are one of the most important road components, requiring routine maintenance for road safety. These structures initially require visual inspection to identify and prioritize maintenance processes based on the damage severity deteriorating the structural strength of the bridge. However, owing to the limitations concerning field inspector resources and budget management efficiency, automation technology is introduced to support field inspectors for developing faster defect detection and inspection processes with high accuracy on a large scale. It has advantages in utilizing the limited number of field inspectors more efficiently and preventing the development of defect severity, which requires more time and budget resources for maintenance. This paper describes the complete development of a deep learning-based visual defect-inspection system for reinforced concrete bridge substructures. The system consists of four main components. The first part involves the image acquisition. The second part aims to detect images with defects using a modified ResNet-50 CNN, improved from our previous research. The third part is defect classification, where different types of defects, such as cracking, erosion, honeycomb, scaling, and spalling, are classified using a previous modified ResNet-50. The last part considers the severity prediction using ANN. As a result, the inspection accuracy rate for defect detection, classification accuracy, and severity prediction are 90.4%, 81%, and 78%, respectively. The promising results of this study were acceptable by Thailand’s Department of Highways for practical use. Finally, this study could support Thailand’s Department of Highways and surrounding construction industries as a standard strategy toward digital transformation.

钢筋混凝土桥梁下部结构是最重要的道路构件之一，需要定期维护以确保道路安全。这些结构最初需要目视检查，以根据损坏桥梁结构强度的严重程度来识别和确定维护过程的优先级。然而，由于现场检查员资源和预算管理效率的限制，引入了自动化技术来支持现场检查员大规模开发更快速、高精度的缺陷检测和检查流程。它在更有效地利用有限数量的现场检查员和防止缺陷严重程度的发展方面具有优势，这需要更多的时间和预算资源进行维护。本文描述了用于钢筋混凝土桥梁子结构的基于深度学习的视觉缺陷检查系统的完整开发。该系统由四个主要部分组成。第一部分涉及图像采集。第二部分旨在使用改进后的 ResNet-50 CNN 检测有缺陷的图像，该 CNN 从我们之前的研究中得到改进。第三部分是缺陷分类，其中使用先前修改的 ResNet-50 对不同类型的缺陷（例如开裂、侵蚀、蜂窝、结垢和剥落）进行分类。最后一部分考虑了使用 ANN 的严重性预测。结果，缺陷检测、分类准确率和严重性预测的检查准确率分别为 90.4%、81% 和 78%。这项研究的有希望的结果被泰国公路部接受并用于实际应用。最后，这项研究可以支持泰国公路部和周边建筑行业作为数字化转型的标准战略。