Robust three-stage deep learning and image processing framework for automated loose bolt detection in complex environments

Vision-based bolt looseness detection is critical for infrastructure safety, yet current methods struggle with bolts of diverse scales, types, and viewing angles in complex environments. This research addresses the challenge of achieving accurate looseness identification for multi-type bolts under such conditions. A three-stage framework is proposed that decouples the task into bolt localization using improved YOLOv8, fine-grained classification via the lightweight RepViT network, and multi-strategy looseness recognition of image processing and deep learning. The method achieves high accuracy and efficiency across all stages, with localization recall at 96.1%, classification accuracy at 98.4%, and final looseness identification accuracy up to 94.5%. This research will advance the application of machine vision in defect identification and intelligent maintenance within the construction sector. The phased methodology may similarly be applied to defect detection in other infrastructure domains, and extended to develop end-to-end integrated systems.