Delamination detection in concrete decks using numerical simulation and UAV-based infrared thermography with deep learning

Abstract

The potential of concrete bridge delamination detection using infrared thermography (IRT) has grown with technological advancements. However, most current studies require an external input (subjective threshold), reducing the detection's objectivity and accuracy. Deep learning enables automation and streamlines data processing, potentially enhancing accuracy. Yet, data scarcity poses a challenge to deep learning applications, hindering their performance. This paper aims to develop a deep learning approach using supervised learning object detection models with extended data from real and simulated images. The numerical simulation image supplementation seeks to eliminate the limited data barrier by creating a comprehensive dataset, potentially improving model performance and robustness. Mask R-CNN and YOLOv5 were tested across various training data and model parameter combinations to develop an optimal detection model. Lastly, when tested, the model showed a remarkable ability to detect delamination of varying properties accurately compared to currently employed IRT techniques.