Semantic Segmentation Models for Crack Detection: Using Shelled Unmanned Aerial Vehicle Imagery

Infrastructures are omnipresent today, with some structures weakening over the years due to natural disasters and aging. Periodic structural health monitoring is essential to keep the health status of the infrastructure in check and avoid major risks and accidents. One important indicator of the infrastructure’s health status is the presence of surface cracks. Meanwhile, doing manual inspection on possible cracks has multiple setbacks, such as risky climbing, time-consuming, and limited attainability. Furthermore, this method poses additional risk to the inspectors, especially in high-rise infrastructures. Thus, the adoption of unmanned aerial vehicles with a passively rotating shell allows close-proximity inspection of surface cracks in infrastructures. In this paper, we present a comparison between existing semantic segmentation models such as the U-Net, LinkNet, FPN, and UNet++ trained for crack detection using shelled UAV imagery. This paper also introduces a new curated dataset, CrackUAS, which contains three types of images: (i) images of unobstructed surface cracks, (ii) images taken from the perspective of the shelled UAV that shows portions of the passively rotating shell, and (iii) synthetic images of surface cracks with shell obstruction. Experimental results of the study demonstrated that the segmentation models could learn crack segmentation from our CrackUAS dataset with state-of-the-art performance based on the precision, recall, and F1 score metrics. The results show that the U-Net and U-Net++ models are neck and neck in terms of performance. For applications where crack detection is used in real-time, the experiments show that FPN is the best choice.