License Plate Character Recognition with Lightweight Convolutional Neural Networks

Character recognition is a crucial step in license plate (LP) recognition. Traditional methods based on template matching and human feature extraction are not robust enough. The Convolutional Neural Networks (CNN) based algorithm can automatically extract the character features through the deep network and significantly improve the correct recognition rate. This paper investigated data augmentation and a lightweight CNN-based algorithm for character recognition. The lightweight network is proposed to enhance character feature extraction by optimizing the convolutional network. In addition, various data enhancement methods are applied to increase the number of data samples to improve the robustness of recognition and enhance the model generalization ability. The experimental results showed that the character recognition accuracy of the proposed approach reached 99.8% with high precision, recall, and specificity. The recognition time for a single character was 6.9 ms on CPU and 0.46 ms on GPU, which meets the demand for real-time LP recognition and has specific reference significance in intelligent transportation.