DSE-YOLO: An Improved Road Damage Detection Model Based on YOLOv8

Road damage detection is critical for ensuring the longevity and safety of road infrastructure. However, this task is challenged by factors such as poor image contrast, complex and irregular damage distributions, and minor damage, all of which compromise detection accuracy. To address these issues, we propose DSE-YOLO, an improved road damage detection model built upon YOLOv8. First, we integrate the C2f-DWR module, which enhances multi-scale feature extraction. Next, we incorporate soft nearest neighbor interpolation (SNI) in the feature fusion layer to resolve feature misalignment issues. Additionally, an efficient multi-scale attention (EMA) module embedded in the detection head enhances detection performance in complex scenarios by aggregating pixel-level features through dimensional interaction. Finally, we introduce the Ins-IoU loss function, which prioritizes the shape of bounding boxes and incorporates auxiliary boxes, thereby accelerating convergence and improving regression accuracy. Extensive experiments on two open-source datasets, SVRDD and CNRDD, show that our method outperforms the baseline, with improvements of 4.4% and 5.5% in mAP@50 and mAP@50-95 on the SVRDD dataset, and 4.4% and 4.5% in mAP@50 and F1-Score on the CNRDD dataset. These results highlight the effectiveness of our approach compared to state-of-the-art methods.