Pavement Distress Detection Model Based on Improved YOLOv10-CSS

To address the issues in existing pavement distress detection models, such as weak feature extraction capability, imbalance between detection accuracy and model efficiency, and dimensional reduction in channel attention, we propose an improved pavement distress detection model based on YOLOv10. The model introduces an enhanced feature map upsampling mechanism (CARAFE), a spatial-channel synergistic attention module (SCSA), and an SioU loss function, named YOLOv10-CSS. The CARAFE mechanism is integrated into the neck network of YOLOv10n to replace the original upsampling module, using the content of input features to guide the upsampling process, which improves the accuracy and efficiency of feature reconstruction and enhances the model's ability to extract relevant features. Additionally, the SCSA attention mechanism is added to the neck network to leverage multi-level semantic information, improving performance in visual tasks and boosting detection accuracy while maintaining a lightweight model. Moreover, replacing the bounding box regression loss with SIoU speeds up the model's regression, enhances robustness, and improves detection performance. Experimental results on the publicly available road damage detection dataset RDD2022 show that the improved model achieves a 1.3 % increase in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP} {@} 0.5$</tex> and a 0.9 % increase in [email protected] compared to the original model, with only a 5 % increase in parameters. The model's effectiveness and superiority are further validated, providing technical support for global road safety monitoring and prevention.