LBN-YOLO: A Lightweight Road Damage Detection Model Based on Multiscale Contextual Feature Extraction and Fusion

Detecting and classifying road damage are crucial for road maintenance. To address the limitations of existing road damage detection methods, including insufficient fine-grained contextual feature extraction and complex models unsuitable for deployment, this paper proposes a lightweight backbone and neck road damage detection model named LBN-YOLO. First, the backbone and neck of the original model are improved to be lightweight, and the C2f-dilation wise residual (C2f-DWR) module is integrated in the backbone to extract multiscale contextual information. Second, a simplified bidirectional feature pyramid network is employed in the neck structure to optimize the feature fusion network, reducing the number of parameters and simplifying the model complexity. Finally, a dynamic head with self-attention is introduced to enhance the sensing capability of the detection head, thus improving the precision of detecting occluded small objects. The proposed model’s detection ability is evaluated using a custom road damage dataset. The experimental results demonstrate that our proposed LBN-YOLO model achieves superior performance compared with the YOLOv8n model, with an increase of 4.1% in [email protected] and a 5.2% enhancement in precision, outperforming other detection models. In addition, the model is evaluated on two public datasets, showing improved detection performance compared with the original model, demonstrating strong generalization capabilities. Code and dataset are available at https://github.com/gzNiuadc/Road-crack-dataset.