Deep Learning-Based Multi-Defect Detection On Bridge Surfaces Using An Improved Yolov8 Model

Abstract

Bridge structures are vital components of transportation infrastructure, and their safety
depends on early detection of structural damage such as cracks, corrosion, and spalling.
Conventional inspection methods rely on manual visual examination, which is slow, laborintensive,
and prone to human error. Recent computer vision approaches attempt automated
inspection; however, most existing models detect only a single type of defect and perform
poorly when multiple defects appear simultaneously on the same surface. This work proposes
an improved deep learning framework based on the YOLOv8-CBAM-Wise-IoU model for
detecting multiple bridge surface defects. Bridge images are captured using cameras or drones
and preprocessed to improve quality. The YOLOv8 backbone extracts visual features, while
the Convolutional Block Attention Module (CBAM) focuses on important defect regions and
suppresses background noise. The Wise-IoU loss function enhances bounding-box
localization, improving detection of overlapping and irregular defects. Experimental
evaluation demonstrates strong performance of the proposed model, achieving an accuracy of
97.9%, a recall rate of 76%, an F1-score of 58%, and a mean Average Precision (mAP50) of
55.4%. The system supports real-time inspection and provides visual outputs to assist
maintenance planning. By enabling early identification of structural damage, the proposed
approach reduces inspection cost, increases reliability, and enhances bridge safety.