Fuzzy License Plate Recognition Based on Recurrent Convolutional Neural Network and Generative Adversarial Network

Jiangjiang Li and Lijuan Feng1*

Abstract

Fuzzy license plate recognition is a difficult problem in the field of license plate recognition. In view of the difficulties in collecting fuzzy license plate images, the large license plate recognition algorithm model, and the shortcomings of mobile or embedded devices, this paper proposes a lightweight fuzzy license plate recognition method, which uses deep convolutional generation adversarial network to generate fuzzy license plate images. The algorithm mainly consists of two parts, namely fuzzy license plate image generation based on optimized convolutional generation adversarial network and lightweight license plate recognition based on deep separable convolutional network and bidirectional long short-term memory. Combined with the lightweight license plate recognition model with depth separable convolution, the recognition rate is comparable to that of the license plate recognition method based on the standard convolutional recurrent neural network (CRNN) after being improved by the images generated in this paper. However, the size and recognition speed of the model are better than that of the standard CRNN model, which is used to solve the problem that it is difficult to collect fuzzy license plates in the real scene, and improves the recognition accuracy of the algorithm while improving the deployment generalization ability. Using generative adversarial network to generate images can effectively solve the problem of insufficient fuzzy license plate image samples. Combined with the depth separable convolution of lightweight license plate recognition model, it has good recognition accuracy and better deployment generalization ability.