PSI - Issue 64

Konrad Bergmeister et al. / Procedia Structural Integrity 64 (2024) 14–20 Konrad Bergmeister / Structural Integrity Procedia 00 (2019) 000 – 000

17

4

Fig. 1. Example of equipment for high-resolution image acquisition (DJI Matrice 350RTK equipped with ZENMUSE H20 camera).

2.2. Image recognition and machine vision Image recognition and machine vision are increasingly employed for various applications (e.g., face recognition). Therefore, applying such straightforward technology to detect structural defects from UAV-obtained images automatically has been widely investigated over the last decade, utilizing various machine learning techniques, including deep learning. Koch et al. (2014) relied on machine vision to inspect large concrete structures and reconstruct three-dimensional digital representations incorporating their physical and functional characteristics, i.e., Building Information Models (BIM), while summarising the achievements and open challenges in vision-based detection of cracking and spalling. This specific field has experienced impressive growth over the last years. Liang (2019) employed an AI-based deep learning architecture to develop an approach that consists of three sequential levels and automatically conduct system level failure classification, component-level bridge column detection, and local damage-level damage localization. The state-of-the-art review of Sabato et al. (2023) studied noncontact techniques used for structural health monitoring and using UAV-obtained images for point cloud reconstruction with the integration of AI methods. Their review surveyed one hundred thirty-five papers and concluded that these methods have great potential. However, their performance under poor lighting conditions and extreme environments still needs to be verified. The current work aims to assist with the time-consuming post-processing of the accumulated data volume of all digital images. Therefore, appropriate AI-based algorithms are developed to identify different types of damage. To this end, a database is created with humanly annotated images of defects (i.e., damages) from different bridge components. This labelled database is then used to train and test the AI models to obtain defect annotations on UAV obtained images automatically. The defects automatically identified and annotated are cracks, corrosion of steel reinforcement induced defects, spalling, cracks with precipitation, algae and net-cracks. Net-cracks are caused by inadequate concrete curing and are characterized by the formation of random cracks with a small crack width (w<0,1 mm). As the resulting crack pattern looks like a net, they are called net-cracks. As an example, Fig. 2 shows the UAV-obtained image of the base of a bridge pier (on the left) and the AI-enhanced images with annotations of two defects, cracking (in red) and spalling (in blue) (on the right). The automatic defect annotation also contains essential information on the exact location within the structure and the specific geometry of the defect (e.g., crack length, location, area, etc.).