Artificial Intelligence‐Driven Multi‐Class Road Damage Detection Using Attention‐Enhanced YOLOv8 and Weighted Boxes Fusion

ABSTRACT Accurate road damage detection is essential for pavement maintenance, traffic safety, and intelligent transportation systems. However, road damage targets often exhibit large variations in shape, scale, texture, and boundary clarity, making detection difficult under complex pavement backgrounds. To address these challenges, this study proposes an Efficient Multi‐Scale Attention Fusion YOLOv8 framework, termed EF‐YOLOv8, for multi‐class road damage detection. The proposed framework combines an Efficient Multi‐Scale Attention (EMA)‐enhanced YOLOv8 branch, a Faster Region‐based Convolutional Neural Network (Faster R‐CNN) branch, and a Weighted Boxes Fusion (WBF) module. YOLOv8 is used as the efficient baseline detector, while the EMA module is introduced to strengthen multi‐scale road damage feature representation and suppress irrelevant background information. Faster R‐CNN is further incorporated as a complementary two‐stage detector to provide additional localization information. The final detection results are generated through WBF, in which overlapping predictions from different detectors are integrated according to confidence scores and model weights. Experiments were conducted on the Japan subset of the Road Damage Dataset 2022 (RDD2022), including four road damage categories: longitudinal cracks, transverse cracks, alligator cracks, and potholes. An additional cross‐dataset validation was further conducted on the India and Czech Republic subsets of RDD2020 to examine the external generalization ability of the proposed framework under different regional pavement and imaging conditions. The results show that EF‐YOLOv8 achieves the best overall performance among all compared models, with a Precision of 0.6359, a Recall of 0.5729, an mAP50 of 0.5870, and an mAP50‐95 of 0.2584. Class‐wise results further indicate that the proposed framework improves the detection stability of different road damage categories. Although the fusion framework increases inference time, it provides more reliable detection results and is suitable for road inspection scenarios where detection accuracy is prioritized. The main contribution of EF‐YOLOv8 lies in its task‐oriented asymmetric fusion design, in which EMA‐YOLOv8 serves as the primary feature‐enhanced detector and Faster R‐CNN is used as an auxiliary detector to provide complementary candidate boxes. The WBF module then integrates these predictions with validation‐optimized fusion parameters, allowing the framework to improve detection stability beyond a direct combination of existing modules.