Abstract
BACKGROUND: With the rapid growth of the number of cars and the increasing complexity of urban transportation, it is particularly important to achieve high-accuracy license plate recognition in complex scenarios. However, since license plate recognition models are mostly deployed on embedded devices with limited computational resources, designing a lightweight and accurate model has become an urgent problem in the field of license plate recognition. METHODS: This study proposes an improved license plate recognition algorithm. We use the License Plate Recognition Network (LPRNet) as the base model. To enhance its accuracy, we incorporate graphically supervised signals for assisted training. This approach refines the training process, yielding a model that is both lightweight and highly accurate. An auxiliary training branch is added, utilizing these graphical signals to guide the model in learning improved image features. RESULTS: Experiments show that compared with LPRNet, this study improves the accuracy in all test sets of the Chinese City Parking Dataset (CCPD) dataset, where the average accuracy is improved by 5.86%, the maximum accuracy by 10.9%, the average character precision by 2.1%, and the average recall by 6.9%, indicating that this study can achieve higher accuracy while keeping it lightweight. This study also provides new ideas for other deep learning image recognition tasks.