SE-YOLOv8: a model for seaside pavement crack detection in remote sensing images

Coastal roads are highly prone to cracking due to prolonged exposure to seawater erosion, tidal fluctuations, and wind-blown sand, posing significant threats to traffic safety and infrastructure durability. To achieve accurate detection of cracks on coastal roads, we propose an SE-YOLOv8 model for coastal pavement crack detection in remote sensing images, based on the YOLOv8 algorithm and incorporating an improved Swin Transformer module and an efficient multi-scale attention (EMA) attention mechanism module. First, we incorporate learnable residual scaling factors into the original Swin Transformer module, enabling the network to adaptively select processing strategies for different crack features, thereby capturing crack details more precisely and enhancing overall detection accuracy and generalization ability. Furthermore, we replace the C2f module in the original YOLOv8 with the EMA attention mechanism, allowing the model to dynamically focus on the most relevant crack features at different levels, thereby enhancing its robustness to environmental noise. Finally, to adapt to the task of crack detection in complex environments, we introduce affine transformation operations with randomized parameters to simulate crack variations that may occur under different conditions, effectively increasing the diversity of the dataset. Experimental results demonstrate that the proposed SE-YOLOv8 model achieves higher detection accuracy on the pavement dataset. Compared with faster R-CNN, YOLOv5, YOLOv8s, and YOLOv10 models, the SE-YOLOv8 model improves mAP@0.5 by 36.8%, 5.6%, 3.9%, and 9%, respectively. These results demonstrate the advantages of the proposed SE-YOLOv8 model in pavement crack detection using marine remote sensing images, showcasing its strong potential for practical applications.