Research on License Plate Recognition Method Based on LPRNet and Improved YOLOv8

Aiming at the problems of low recognition efficiency, weak generalization ability, and low recognition accuracy of existing license plate recognition technology, an end-to-end license plate detection and recognition model was proposed for license plate detection and recognition in a high- precision complex environment. First of all, the license plate location model uses YOLOv8, in which the Biformer attention mechanism is introduced to enhance the attention of small targets in the license plate, and the traditional convolution is replaced by the deep separated convolution (DSC), which reduces the number of parameters and improves the calculation speed, thus producing a more lightweight model. The license plate recognition network uses LPRNet, which is a lightweight network proposed specifically for license plate recognition. Its model size is only about 1.5M. Finally, this paper presents a large number of experimental results on the Chinese city license plate dataset (CCPD2019) and the selfie dataset. The results show that the mAP50 of the license plate location model based on the improved YOLOv8 proposed in this paper is 0.994, and the MAP50:95 is 0.922. In addition, the mAP value of the LPRNet license plate recognition model also reached 0.896. The overall recognition accuracy of the model is 98.05%, which is obviously better than the traditional license plate recognition technology. It performs well in complex environments, maintaining high accuracy while remaining lightweight, and is easy to deploy on embedded edge devices.