CDD-TR: Automated concrete defect investigation using an improved deformable transformers

Abstract

Public infrastructures, such as bridges, dams, and buildings, play a key role in urban development. Structural inspection by visually monitoring and inspecting the structures for defects has become increasingly vital to prevent structural deterioration. However, previously, the structural inspection was primarily carried out manually, which was time-consuming, error-prone, and tedious. Therefore, this study proposes an efficient concrete defect detection system based on a transformer model. Four primary contributions are (i) a novel defect detection framework motivated by the deformable transformers (Deformable DETR); (ii) the use of a big concrete defect dataset containing four common defect types; (iii) multiple modules are introduced to the original Deformable DETR model and help the model achieve better performance; and (iv) visualization of the model’s deformable attention weights to show the model effectiveness in detecting and localizing defects. The framework outperforms previous state-of-the-art object detection networks and obtains the mean Average Precision (mAP) of 63.8%.