Efficient CNN architecture for traffic sign recognition in visually challenging weather conditions

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This study proposes an efficient convolutional neural network for traffic sign classification in visually challenging weather, including fog and snow. The model is trained on the German Traffic Sign Recognition Benchmark and evaluated on both weather-augmented images and real-world weather-degraded scenes. Adverse conditions are modeled using snow and fog simulations, and the model’s generalizability is further validated on real datasets captured in foggy and snowy environments. We compare the proposed network with state-of-the-art architectures (VGG16, A-DCNN, VGG19, E-MobileViT, and ResNetV2-50) and show comparable or better accuracy, precision, recall, and F1-score while using far fewer trainable parameters (188,235). The model remains robust across visibility levels without complex preprocessing or a large-scale backbone. It achieves 99.47% accuracy in clear conditions and exceeds 94% under adverse weather, supporting suitability for embedded and real-time deployment. Extensive experiments show that joint training on original and weather-augmented images improves recognition in adverse conditions while preserving performance on clear images with a single model. Overall, the proposed approach offers a resource-efficient and accurate solution for intelligent transportation systems operating in visually degraded environments.