Improved YOLOv4 based on feature enhancement for vague traffic sign detection

Due to the interference of external factors, traffic signs are easily blurred in the image, which greatly reduces the detection accuracy of traffic signs. At the same time, considering that most of the application scenarios of traffic sign detection require high real-time performance, a vague traffic sign detection algorithm based on road feature enhancement and improved YOLOv4 is proposed. First, to suppress the cluttered background information in the traffic sign images and effectively capture the traffic sign features, coordinate attention is embedded in the feature extraction part of YOLOv4. Then, YOLOv4 adopts max pooling for downsampling, ignoring the loss of shift and other variability, which easily leads to unstable feature extraction of fuzzy traffic signs. In order to extract image features more effectively, BlurPool is used for downsampling to more effectively retain the features of blurred traffic signs. Finally, DUpsampling is used for upsampling in the feature fusion stage to establish the correlation between the newly inserted pixels and the original pixels. The experimental results show that the size and parameters of the improved YOLOv4 model are reduced by 10.18% and 11.32%, respectively, compared with the original algorithm, and the FPS and mAP are increased by 2.02 and 1.34, respectively. By comparing with YOLOv3_SPP, SSD, Faster RCNN and other traffic sign detection algorithms, the improved YOLOv4 outperforms these algorithms. Through the input image test, the improved algorithm in this paper has better performance for fuzzy traffic sign detection in real scenes.