Automatic detection of subsurface defects in concrete structures using state-of-the-art deep learning-based object detectors on the infrared dataset

Abstract

Coupling infrared thermography (IRT) with deep learning is a potential approach to address the limitations of conventional IRT-based concrete delamination identification. However, it remains under-explored, mainly due to the scarcity of infrared data. This study examined numerous advanced object detectors, including Faster R-CNN, SSD, and the variants of YOLO versions (YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLO-NAS, and YOLOv9) to comprehensively evaluate their applicability on a self-prepared infrared dataset comprising 12,000 images of artificial defects. The results demonstrated that most of the selected models worked efficiently on the IRT dataset with precision, recall, F1-score, and AP₅₀ greater than 91 % and real-time speeds higher than 45 FPS. The effectiveness of each model was elucidated, and the most suitable frameworks for delamination detection systems were proposed. Interestingly, newer versions of the selected YOLO models failed to surpass their predecessors. Moreover, larger variants did not outperform smaller ones within each YOLO version. Additionally, the effects of various IRT experimental conditions, image processing techniques, and data augmentation strategies on the model’s performance were thoroughly investigated. Consequently, the capabilities and limitations of the automated approaches presented in this work form a solid foundation for further studies on automatically detecting internal defects in concrete structures using deep learning models and IRT inspection data.