A robust bridge rivet identification method using deep learning and computer vision

摘要

Timely and effective inspection ensures safe operation and optimum resource use for infrastructure maintenance and renewal. Robot advances allow rapid collection of inspection image data. However, distinguishing bridge elements from large amounts of image data is challenging. Rivets are critical elements, joining different profiles into components. However, automatic rivet identification has received little attention. This study proposes a rivet identification method based on computer vision and deep learning. A sustainable training framework is presented to build a robust detector. A novel rivet dataset was collected and annotated from a full-size bridge. YOLOv5 is used to extract features and predicate classifications. The model achieved an 88.9% precision, 90.5% recall, and 90.1% F1 score. The accuracy and robustness were evaluated on another riveted bridge under various operational conditions. The rivet detector generally performs well, achieving 85% or even 95% accuracy in most situations. Out-of-focus and object occlusion have the largest negative effect.