Surface Damage Detection for Bridge Cables Based on YOLOv7

There is a growing demand for automatic bridge cable inspection using intelligent detection technologies to replace manual inspection. However, the application of intelligent inspection systems is hindered due to the lack of effective cable damage detection methods. To resolve this issue, this paper proposes a surface damage detection method for bridge cables based on YOLOv7, a powerful real-time object detector. We introduce the basic algorithm of YOLOv7, and elaborate its training, transfer, and deployment methods for cable damage detection. A dataset is built containing 7548 cable damage images collected from real bridge cables. The results show that this method can effectively detect common bridge cable defects, including scratches, dirt, and abrasions. The proposed method achieved an F1 score of 0.75 and an mAP score of 0.799 on the test set. The incorrect predictions are mainly caused by slight and vague defects that can be either labeled or not labeled. The inference efficiency of the proposed method reaches 20.5 FPS for images with resolution of 640×640 when deployed with NVIDIA Triton. Considering the diversity of bridge cables in real-world scenarios, we have also conducted model transfer analysis to discover how much additional data is needed to transfer a pre-trained model for a new bridge, which provides a valuable reference for applying transfer learning technology on bridge cable inspection. The proposed approach is practically useful for assisting humans in data processing, reducing labor input significantly, and improving the objectivity and consistency of cable inspection.