Research on License Plate Recognition Algorithm Based on YOLOv5 and LPRNet

The traditional license plate recognition (LPR) algorithm is susceptible to the photo-taking conditions such as the weather, illumination and angle. Thus, the deep learning method is introduced to solve this problem. Using YOLOv5 to locate license plates can reduce environmental interference and improve location accuracy. The obtained license plate area is corrected through perspective transformation and then as an input of LPRNet. LPRNet can be used to identify characters without character segmentation, so the disruption of the border is avoided. In this method, we use Multi-Scale Retinex with Color Restore (MSRCR) to enhance pictures. Finally, end-to-end license plate recognition is realized by connecting YOLOv5 and LPRNet models in series. The experimental results show that the accurate localization rate reaches 99.5% with the Intersection over Union (IoU) of 0.5 after image enhancement. The recognition accuracy of characters is 96%. Compared to traditional LPR methods, our method improves the accuracy of license plate location and can directly recognize characters without segmentation.