Abstract: At present, the license plate recognition model based on neural network is large, which is not conducive to deployment in edge computing equipment. In addition, the recognition rate of oblique license plates and mixed color license plates is low. To this end, a lightweight neural network combination model for recognizing tilted mixed color license plates is proposed. The model first uses the improved P-YOLO algorithm based on YOLO algorithm to achieve license plate detection, classification, and localization; Then, an improved N-G-LPRNet algorithm based on LPRNet algorithm is used to recognize license plate characters. Finally, compared with the improved algorithm model, the test results on the CCPD dataset show that the P-YOLO algorithm significantly improves the mAP index for detecting tilted license plates during the license plate detection stage. Combined with character recognition networks, the recognition accuracy for conventional license plates, slightly tilted license plates, and strongly tilted license plates is improved by about 1.3%, 70.7%, and 63.8%, respectively; In the character recognition stage, under the premise of mixed color license plate training, the N-G-LPRNet algorithm improves the recognition rates of blue and green license plates by about 40.65% and 32.26%, respectively; The final P-YOLO-N-G-LPRNet combination model has a comprehensive recognition rate of 98.16%, occupying approximately 5MB of space. It has obvious advantages in high recognition rate and lightweight.

Key words: Deep learning, Target detection, Character recognition, YOLO, LPRNet