ABSTRACT
In 2021, 42% of bridges in the United States were used beyond their design life, incurring a repair cost of approximately $125 billion. To reduce future repair costs of bridges, novel inspection methods have been recently proposed that use unoccupied aerial vehicles to inspect bridge defects, where a computer vision-based model processes the data to assess bridge condition. Although this technology can provide significant advantages in decreasing inspection costs and labor, inspection accuracy relies on the quality of images and the computer vision model. Therefore, this study develops a multi-task deep learning framework that combines object detection and semantic segmentation models to detect important structural details and quantify corrosion level in steel bridges with high precision and speed (i.e., inference time). The framework leverages high-performing computing and high-definition images to minimize resizing effects on model accuracy. For each module, two different deep learning architectures were developed and compared in terms of accuracy and inference time. The results show the efficiency of deep learning models for both tasks. For an intersection over a union threshold of 50%, the object detection models of YOLOv8 and SSD RetinaNet50 achieve mean average precision of 96.6% and 85.2%, respectively. In addition, the segmentation models of YOLOv8 and DeeplabV3+ achieved an F1 score of 61.5% and 28%, respectively.
