In order to overcome the challenges of traffic sign detection and recognition, such as small targets, diverse sizes, difficulty in obtaining relevant feature information, and susceptibility to complex background interference, traffic signs are detected and recognized. a deep learning method for traffic sign detection and recognition is proposed based on the YOLOv5s network. Addressing the issue of current traffic sign recognition algorithms struggling to identify small targets in complex backgrounds, we have integrated a shuffle attention mechanism into the C3 layer at the end of the YOLOv5s backbone network. This integration introduces a traffic sign detection and recognition algorithm that relies on an attention mechanism. This enhances the ability to focus on key areas and effectively eliminates background noise interference. To address the limitations of feature fusion in current object detection algorithms when dealing with traffic signs of varying sizes in images, we propose a weighted feature fusion network algorithm. This algorithm performs weighted fusion of shallow feature maps containing rich semantic information in the backbone network with medium and large target detection layers, enhancing the fusion ability of multi-size features. Experimental results on the traffic sign detection dataset CCTSDB 2021 show that the enhanced algorithm achieved a 0.5 percentage points increase in precision, a 3.6 percentage points increase in recall, and an average precision improvement of 2.8 percentage points compared to the original YOLOv5s method. Additionally, the detection speed reached 123.46 frame/s. Therefore, the proposed algorithm effectively enhances the accuracy of traffic sign detection and recognition while maintaining a original detection speed.

Small object detection  /  Traffic sign recognition  /  YOLOv5s  /  Shuffle attention mechanism  /  Weighted feature fusion