Enhancing swin transformer via ProbAttention and Convolutional Block Attention Module for traffic sign recognition

In autonomous driving systems, traffic sign recognition (TSR) serves as a pivotal component for safeguarding driving safety and boosting operational efficiency. Conventional approaches to TSR, which depend on manual visual assessment, suffer from high labor intensity, excessive time consumption, and a significant propensity for errors. To tackle these drawbacks, this study presents an optimized framework tailored for TSR, which embeds ProbAttention and the Convolutional Block Attention Module (CBAM) into the Swin Transformer architecture. The Swin Transformer is well-recognized for its competence in multi-scale feature extraction, whereas ProbAttention effectively alleviates the quadratic computational complexity inherent in traditional attention mechanisms. Additionally, CBAM enhances the model’s capacity to concentrate on prominent features in traffic sign images by sequentially deducing attention maps across both the channel and spatial dimensions. The proposed model performs 3-way classification (prohibitory, warning, and mandatory) without any detection head or bounding box prediction, making it a pure classification pipeline suitable for real-time deployment. Comprehensive experiments carried out on publicly accessible datasets (CCTSDB2021, GTSRB, and BTSD) demonstrate that the proposed method achieves competitive performance compared to existing algorithms. Not only does this approach improve recognition accuracy over the baseline architecture, but it also lowers computational complexity—rendering it a potentially valuable tool for real-time traffic sign recognition scenarios in autonomous driving systems.