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

Ensuring road safety and efficiency, traffic sign recognition (TSR) is a crucial aspect of autonomous driving systems. Existing TSR methods, however, involve manual inspection, leading to high labor costs, extended processing times, and a high error rate. This research introduces an optimized TSR framework that incorporates ProbAttention and the Convolutional Block Attention Module (CBAM) into the Swin Transformer architecture.

Renowned for its multi-scale feature extraction, Swin Transformer benefits from ProbAttention, which reduces the quadratic computational complexity of traditional attention mechanisms. Additionally, CBAM aids the model in focusing on critical features in traffic sign images by sequentially generating attention maps for both channel and spatial dimensions.

The proposed model is capable of classifying traffic signs into three categories (prohibitory, warning, and mandatory) without the need for detection heads or bounding box predictions, making it a pure classification pipeline ideal for real-time applications. Evaluations conducted on CCTSDB2021, GTSRB, and BTSD datasets reveal that the proposed method offers competitive performance against existing algorithms while also reducing computational complexity.

Consequently, this approach presents a promising solution for real-time traffic sign recognition in autonomous driving environments. This study was supported by the Foundation of Doctoral Scientific Research of Shandong Management University (SDMUD202125).