Vision‐based robust missing and loosened bolt detection for splice plate joints

Summary Vision‐based bolt defect detection methods based on feature changes have been reported. However, the robustness of key feature extraction and bolt detection requires improvement. This paper proposes a robust missing and loose bolt defect detection approach. The key features—reference points for perspective correction and the straight lines of the bolt edges—are extracted from the masks obtained by semantic segmentation models. The true and false bolt discrimination approach based on the mask shape can help improve bolt object detection accuracy. Overlapping between the bolt bounding boxes in the reference and detection images indicates missing bolts. The rotation angles reveal loosened bolts. The proposed approach was tested on fabricated bolted joint specimens and a steel railway bridge. The results suggest that these improvements ensure defect detection accuracy, with a miss rate of only 1% for missing bolt detection. Moreover, a loosened bolt with only 3° rotation is successfully detected. This approach has promising potential applicability in automatically detecting bolt defects in large steel structures.