A Lightweight Traffic Sign Recognition Network for Autonomous Driving: Integrating Dual-Branch Attention and Weather-Robust Augmentation
Main Article Content
Authors
Changsha Medical University
Zhejiang Automotive Engineering Advanced Research Institute
Abstract
Traffic sign recognition is a key perception task in autonomous driving and advanced driver-assistance systems, where accuracy and real-time performance directly affect road safety. In real-world scenes, fog, rain, and low illumination reduce image contrast and introduce noise, degrading deep models, while in-vehicle platforms impose strict limits on model size and computation. We propose CAMSNet, a lightweight traffic sign recognition network. Its architecture uses a lightweight residual backbone based on depthwise separable convolutions and a Dual-branch Attention Calibration Module (DACM), whose parallel channel- and position-attention branches model channel dependencies and spatial location. Their weights are fused elementwise and applied through residual reweighting to enhance sign features adaptively. During training, a weather-robust augmentation strategy based on atmospheric scattering randomly adds synthetic fog, Gaussian noise, and brightness changes. This augmentation is applied only to training data, leaving inference as a lightweight, one-way network. On the German Traffic Sign Recognition Benchmark (GTSRB), CAMSNet achieves 98.19% top-1 accuracy with 0.458 million parameters and 35.88 million multiply-accumulate operations (MACs), exceeding ResNet-18, which has about 24 times as many parameters, and MobileNetV2, which has about five times as many. CAMSNet with weather augmentation (CAMSNet+W) reaches 98.50% accuracy and 96.34% under severe synthetic fog. This is 36.98 percentage points above ResNet-18 and 36.12 points above CAMSNet without weather augmentation. CAMSNet+W also shows the smallest accuracy drops under severe noise and low brightness. Ablation and controlled comparisons indicate that attention improves robustness under degraded conditions while maintaining similar clean-domain accuracy. DACM is competitive with ECA and CBAM and offers a useful balance between parameter count and robustness. In this single run, weather augmentation improved robustness without an observed loss in clean-test accuracy. Robustness conclusions are based on synthetic corruptions; generalization to real adverse-weather scenes remains to be established.
Article Details
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Copyright remains with the author(s). All articles are published under the Creative Commons Attribution–NonCommercial 4.0 International License (CC BY-NC 4.0). This license permits copying, sharing, adaptation, distribution, and reproduction in any medium or format for non-commercial purposes, provided that appropriate credit is given to the original author(s) and the source, a link to the license is provided, and any changes are indicated.
How to Cite
References
[1] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel, “Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition,” Neural Networks, vol. 32, pp. 323–332, 2012.
[2] S. Grigorescu, B. Trasnea, T. Cocias, et al., “A survey of deep learning techniques for autonomous driving,” Journal of Field Robotics, vol. 37, no. 3, pp. 362–386, 2020.
[3] S. G. Narasimhan and S. K. Nayar, “Contrast restoration of weather degraded images,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 6, pp. 713–724, 2003.
[4] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” in Advances in Neural Information Processing Systems. Lake Tahoe: Curran Associates, 2012, pp. 1097–1105.
[5] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016, pp. 770–778.
[6] A. G. Howard, M. Zhu, B. Chen, et al., “MobileNets: Efficient convolutional neural networks for mobile vision applications,” arXiv:1704.04861, 2017.
[7] M. Sandler, A. Howard, M. Zhu, et al., “MobileNetV2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018, pp. 4510–4520.
[8] J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018, pp. 7132–7141.
[9] S. Woo, J. Park, J. Y. Lee, et al., “CBAM: Convolutional block attention module,” in Proceedings of the European Conference on Computer Vision. Munich: Springer, 2018, pp. 3–19.
[10] Q. Wang, B. Wu, P. Zhu, et al., “ECA-Net: Efficient channel attention for deep convolutional neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020, pp. 11534–11542.
[11] Q. Hou, D. Zhou, and J. Feng, “Coordinate attention for efficient mobile network design,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Nashville: IEEE, 2021, pp. 13713–13722.
[12] G. Zeng, Z. Wu, L. Xu, et al., “Efficient vision transformer YOLOv5 for accurate and fast traffic sign detection,” Electronics, vol. 13, no. 5, Art. no. 880, 2024.
[13] Y. Luo, Y. Ci, S. Jiang, et al., “A novel lightweight real-time traffic sign detection method based on an embedded device and YOLOv8,” Journal of Real-Time Image Processing, 2024.
[14] S. Song, X. Ye, and S. Manoharan, “E-MobileViT: A lightweight model for traffic sign recognition,” Industrial Artificial Intelligence, vol. 3, no. 1, Art. no. 3, 2025.
[15] M. S. Devgan, G. Singh, P. Sharma, et al., “Attention-guided lightweight CNN-transformer fusion for real-time traffic sign recognition in adverse environments: HACTNet,” IET Intelligent Transport Systems, 2026, doi: 10.1049/itr2.70167.
[16] H. Gupta, O. Kotlyar, H. Andreasson, et al., “Robust object detection in challenging weather conditions,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. Waikoloa: IEEE, 2024, pp. 7508–7517.
[17] F. Assion, F. Gressner, N. Augustine, et al., “A-BDD: Leveraging data augmentations for safe autonomous driving in adverse weather and lighting,” arXiv:2408.06071, 2024.
[18] D. Hendrycks and T. Dietterich, “Benchmarking neural network robustness to common corruptions and perturbations,” in Proceedings of the 7th International Conference on Learning Representations, New Orleans, 2019.
Article Sidebar
Download
The GTSRB dataset used in this study is publicly available at https://benchmark.ini.rub.de/gtsrb_dataset.html. No new real-world data were collected for this study.
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.