RepFYOLOv5: A Lightweight Re-parameterization YOLOv5 Network with Four Detection Heads

Deep neural networks have received wide applications in many areas in recent years, but the high requirements of computational cost and storage space limit its applications in embedded devices. To deal with this problem, we proposed a lightweight re-parameterization YOLOv5 network with four detection heads, called RepFYOLOv5. First, a structure re-parameterization module RepBlock is proposed to reconstruct the network backbone, which has multiple topology branches during the training process, but in the inference and deployment, it uses the equivalent convolutional module fusion structure to speed up the network inference. In addition, in order to improve the object detection accuracy (especially for the small objects), except for the original three detection heads of YOLOv5, an extra detection head based on larger feature maps is added to the network to improve the detection accuracy for the small objects. Experimental results on the COCO dataset and our own traffic surveillance dataset showed that our proposed lightweight RepFYOLOv5 obtained a better balance among detection accuracy, inference time and parameter amount etc.