Enhanced YOLOv8 for Detecting Multiple Defects on Bridge Surfaces

Abstract

With the advancement of machine vision, numerous models have been created to detect imperfections in bridges. However, the bulk of these models are designed for single defect detection and are not adept at managing cases with concurrent multiple defects. As a result, quickly recognizing the array of defects on bridge surfaces is still a major obstacle. In response to this challenge, the current research introduces the YOLOv8-CBAM-Wise-IoU model, specifically crafted for the detection of seven distinct bridge surface defect categories. This model integrates the CBAM mechanism for focusing attention and the Wise-IoU for calculating loss, with its effectiveness measured by metrics including accuracy, retrieval rate, F1 measure, and mAP50. Rigorous ablation analyses and benchmarking against both single-tier and multi-tier deep learning frameworks were performed to substantiate the models utility. The YOLOv8-CBAM-Wise-IoU model exhibited formidable performance, recording an accuracy rate of 97.9%, a retrieval rate of 76%, an F1 measure of 58%, an mAP50 of 55.4%, and an mAP50-95 of 32.4%. These results outstrip those of standard models and other ablation variations, emphasizing the models ability to boost the precision and robustness of detecting various defect types on bridge surfaces. Code is available at https://github.com/IamSunday/Enhanced-YOLOv8-for-Detecting-Multiple-Defects-on-Bridge-Surfaces.