Road traffic sign recognition based on multiresolution attention mechanism network

With the development of intelligent transportation systems and autonomous driving technology, road traffic sign recognition has become an important research direction to ensure driving safety and path planning. This paper proposes a deep neural network structure based on Multi Resolution Attention Mechanism (MRAM) to address the problem of decreased recognition rates of existing methods under complex lighting, scale changes, and occlusion conditions. The network introduces multi-scale branches in parallel during the feature extraction stage to capture local and global information, and embeds channel attention and spatial attention modules at each scale to enhance semantic response. To improve the robustness of the model to small-scale and fuzzy labels, this paper also proposes a Residual Cross scale Fusion (RCF) strategy based on residual fusion ensemble, and introduces class difficult sample weighted loss in the training stage to alleviate the problem of class imbalance. Experiments on the GTSRB benchmark dataset and self built complex scene dataset show that the proposed MRAM network outperforms commonly used comparison models such as ResNet and FPN in overall accuracy, micro average F1, and recall of small-scale markers. The method proposed in this article ensures high accuracy while meeting the real-time requirements of vehicle deployment in terms of inference delay.