network model and increased the size of the convolution kernels of the first two layers,
which significantly improved the accuracy of model detection. Samples of detection
results produced by evaluating CMnet on CCPD are shown in Fig. 6.
5 Conclusions
Overall, in this paper we propose a new neural network model for single LP detection.
Our proposed model is superior to the state-of-the-art method in terms of accuracy and
efficiency. In the future work, we will implement the bounding box calibration function
for license plate detection and implement license plate number recognition.
Acknowledgements. This work was supported by the National Natural Science Foundation of
China (No.61976032).
Fig. 5. Feature extraction visualization. a The model before optimization. b The optimized
model
Table 3. Accuracy after model optimization
Model
Accuracy (%)
CMnet-7layer-5size
96.9
CMnet-5layer-3size
97.1
CMnet-7layer-5size
96.7
CMnet-5layer-7size
98.8
CMnet: A Compact Model for License Plate Detection
581