End-to-End License Plate Recognition In Complex Environments Via Adversarial Generation Of Training Examples

License plate ID is one of the key components of various modern intelligent transportation system building requirements.At present,most license plate recognition algorithms are only effective in certain conventional scenes.However,for some uncertain factors in the actual environment,such as the influence of smog,rain and snow weather,vehicle high-speed motion causes the collection of license plates to be blurred,the darkness or backlit of the shooting environment,the monitoring device has low pixel,tilted position,which causes the license plate to be blurred,tilted,and vehicle-related criminal activities such as license plate characters smearing or partial occlusion,etc.The accuracy of the license plate recognition system is still not high.At the same time,there are many types of license plates,such as new energy license plates(8 characters),Hong Kong and Macao license plates,etc.Most of the algorithms only support certain types of license plate recognition,which is difficult to meet the actual application requirements.Although the license plates recognition system based on deep learning has improved the license plate recognition performance to some extent in recent years,there are still some shortcomings in this type of algorithm,such as relying on a large amount of labeled data,requiring character segmentation or label alignment.Therefore,in view of the above problems,this paper is based on deep learning method,and the main research content of this paper can be summarized as the following three points:(1)A good depth model relies on numerous of real-labeled samples and the cost of manual labeling is expensive.In this paper,a new method of data generation and data augmentation strategy is proposed which based on CycleGAN with a “feature consistency” loss function.Using only a few labeled license plate images,large quantity of realistic samples can be created by our model GAN-Plate.(2)In order to improve the robustness of license plate recognition in complex environment,we design a new end-to-end license plate recognition model Incep-PlateNet,which including license plate self-correction module,license plate feature extraction module,etc.It’s an end-to-end model which is free of pre-segmentation or label alignment and supporting multiple types of license plate recognition.(3)Considering the computing performance of the actual deployment of terminal equipment and real-time demand for license plate recognition,this paper follows the model miniaturization design concept when designing the depth model and in the model engineering application deployment phase,we use both of model quantification and model pruning.The model is compressed to improve the speed of the terminal using while ensuring the recognition accuracy of the model.Experiments show that the proposed strategy,adversarial generation of Training Examples,can solve the problem of fewer training samples.At the same time,compared with no data generation,the end-to-end model Incep-PlateNet can achieve 98.87% accuracy in regular environment,which increased by 4.18%.And in complex environment the recognition accuracy can reach 90.06%,which increased by 8.54%.All recognition accuracy better than the state-of-the-art methods.For the actual deployment of the model,the model is only 2.3M and only takes 12-16 ms to test a single license plate recognition on the CPU: Core i7-6700 K Skylake,RAM: 4G.