A YOLOv4 Model with FPN for Service Plates Detection

Intelligent service plates detection plays an important role in smart city. For example, intelligent service plates detection can realize more efficient and accurate self-service charges for restaurants, and saves labor costs, additionally. This paper constructs a service plates detection dataset. On this basis, Faster R-CNN and MobilenetV3 are firstly leveraged to conduct service plates detection. In addition, authors propose an intelligent service plate detection method based on FPN (Feature Pyramid Network) + YOLOv4 network. Specifically, YOLOv4 is utilized to extract initial image features, and a variant of FPN network for YOLOv4 backbone is designed to aggregate multi-granularity image features. To boost the training efficiency, transfer learning algorithm is also introduced into our approach. Through the corporation of FPN + YOLOv4 framework and transfer learning algorithm, they realize the accurate and fast intelligent service plate detection. The experimental results show that: compared with MobilenetV3 model and faster R-CNN, our method achieves great improvement. In the future, authors will improve the network to achieve higher accuracy and faster calculation speed, and expand our data set to more realistic scene images.