Improving YOLOv5 for Traffic Sign Detection

With the development of deep learning models, current traffic sign detection models often suffer from large size and numerous parameters. In order to balance the accuracy, size, and speed of models within limited hardware resources, a lightweight improvement of the YOLOv5 traffic sign detection model is proposed. This model restructures the YOLOv5 network using the GhostNet module, introducing smaller volume modules like C3Ghost and GhostConv. Additionally, it incorporates an ECA channel attention mechanism and replaces the loss function from CIOU to SIOU, enhancing the model's focus on detection targets and improving attention to small objects. Experimental validation on the CCTSDB2021 dataset demonstrates that compared to the YOLOv5 model, this improved model reduces parameter count by 45.5% and model size by 44.4%, emphasizing more accurate detection of real targets. This indicates that the model is more suitable for deployment on embedded platforms than the original model.