Research On Small Sample License Plate Recognition Method Based On Sample Expansion

Vehicle license plate detection and recognition is an important part of intelligent transportation system.The problems of license plate detection and recognition based on traditional methods are weak model generalization ability and low recognition accuracy.Due to the development of computer computing power,license plate detection and recognition algorithms based on deep learning has been widely used.However,the biggest problem of deep learning is that it relies heavily on training samples,and the accuracy of deep neural network often decreases when the sample number of dataset is too small or the environment is changeable,namely,the problem of small samples.At present,there is a lack of some types of license plates（such as new energy vehicle license plates）in the public license plate dataset in China,and the detection and recognition effect based on deep learning is poor.In order to solve this problem,this thesis studies the small sample problem of license plate dataset as follows.（1）A license plate generation model based on generative adversarial network is proposed to expand the scarce samples in license plate dataset.The model is mainly composed of generative network and discriminant network.The generative network generates synthetic samples which are very similar to real samples by simulating the distribution of training samples.The discriminator network needs to discriminate the authenticity of the input sample image.Through adversarial training,the fitting distribution of the generated network is closer to the real sample distribution and the accuracy of the discriminant network is improved.When the accuracy of the discriminant network is stable at about 0.5,the model converges,and the distribution of the synthetic sample is consistent with that of the real sample.In this thesis,a license plate generation model is proposed to effectively expand the scarce samples in the license plate dataset.（2）A method to evaluate the validity of synthetic sample is proposed.The real sample template and the synthetic samples template are extracted and the gray histogram is calculated respectively.The similarity between synthetic sample distribution and real sample distribution is measured by calculating the similarity of gray histogram.After calculation,the Bhattacharyya coefficient and the cosine similarity between the synthetic sample and the real sample are 0.851 and 0.9123,which are far higher than the results before and after the real image is processed in different ways.It proves that the synthetic sample obtained by the license plate generation model proposed in this thesis is effective.（3）A cascade framework of license plate detection and recognition based on small samples is designed to realize the end-to-end detection and recognition of license plates.In this framework,the license plate detection model,the proposed license plate generation model and the license plate recognition model are cascaded together to realize a whole set of solutions to the problem of small samples of license plates.The license plate detection model is built based on one-stage algorithm to achieve fast and efficient license plate detection.The license plate recognition model adopts the idea of not based on character segmentation and introduces the CTC loss function to realize the recognition of indefinite length license plate.Through the expansion of scarce samples by the license plate generation model in the cascaded framework,the recognition accuracy of new energy vehicle license plate is improved by 22.6% and the overall recognition accuracy of blue license plate is improved by 1.1%.