A Lightweight Road Damage Detection Model based on Structural Improvement and Knowledge Distillation

Xiaohui Shi

doi:10.6919/ICJE.202603_12(3).0016

Authors

Xiaohui Shi

DOI:

https://doi.org/10.6919/ICJE.202603_12(3).0016

Keywords:

Road Damage Detection; Lightweight Model; Knowledge Distillation; YOLO.

Abstract

To address the problem of large model parameters in existing high-precision road damage detection models, this paper proposes a lightweight detector called YOLO-LWD based on structural optimization and knowledge distillation. First, standard convolutions in the backbone are replaced with GhostConv, an efficient channel attention (ECA) module is introduced into the neck, and the detection head is replaced with DyHead, with its channel number reduced from 512 to 256. Then, a hybrid knowledge distillation method is adopted to transfer soft labels from the output layer and feature knowledge from intermediate layers of a teacher model to the student model. Experiments on the Aug-RDD dataset show that YOLO-LWD achieves 78.4% mAP@0.5 with only 8.1 MB parameters and 15.8 GFLOPs, which outperforms current mainstream models in both model complexity and detection accuracy.

Downloads

Download data is not yet available.

References

[1] Yang, X., Zhang, J., Liu, W., Jing, J., Zheng, H., & Xu, W. (2024). Automation in road distress detection, diagnosis and treatment. Journal of Road Engineering, 4, 1–26.

[2] Manjusha, M., & Sunitha, V. (2025). Optimizing YOLO models for high-accuracy automated detection and classification of road surface distresses. Innovative Infrastructure Solutions, 10, 381.

[3] Zhang, Y., & Wang, Y. (2020). Machine learning for pavement condition assessment: A review. Journal of Transportation Engineering.

[4] Botezatu, A.-P., Burlacu, A., & Orhei, C. (2024). A review of deep learning advancements in road analysis for autonomous driving. Applied Sciences, 14(11), 4705.

[5] Howard A G, Zhu M, Chen B, et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications[EB/OL]. arXiv, 2017. https://arxiv.org/abs/1704.04861.

[6] Tan M, Le Q V. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks[C]//International Conference on Machine Learning. PMLR, 2019: 6105-6114.

[7] Chen F H, Li S L, Han J L, et al. Review of lightweight deep convolutional neural networks[J]. Archives of Computational Methods in Engineering, 2024, 31(4): 1915-1937.

[8] Golizadeh, M., et al. (2025). Architectural insights into knowledge distillation for object detection: A comprehensive review. arXiv preprint arXiv:2508.03317.