| It plays an unreplaceable role for the License plate recognition in establishing a long-term intelligent transportation system and improving the efficiency of traffic management.However,traditional license plate recognition algorithms using artificially constructed feature extractors have shortcomings such as poor anti-interference ability and poor versatility.Because of the rapid development of artificial intelligence,the realization of license plate recognition through deep learning has become a reality.This paper researches for the methods of license plate location,character correction and character recognition in various scenarios based on Convolutional Neural Network(CNN).The specific content and work are as follows.1.In the link of the license plate positioning,for the case where the license plate is small in the image;this paper uses regional coarse positioning and fine positioning to achieve the precise location of the license plate.Firstly,based on the CNN model of typical Faceboxe,the network was designed by optimizing the fast digestion layer,adding Inception V3 layer and modifying the activation function.An improved Faceboxes model based on network architecture enables rough license plate location and filters out most of the non-license plate areas..Secondly,in order to design a license plate fine positioning algorithm Retina LP with small space complexity,by replacing the backbone structure in the Retinaface model with the Mobilenet V3 small model and optimizing the FPN module,to enhance the backbone network’s ability to learn features and network generalization capabilities,and finally realize the license plate fine positioning.Experiments show that the detection effect of the improved Faceboxes model or Retina LP model proposed in this paper is better than the original algorithm.The accuracy of the license plate location method based on the cascading coarse and fine localization models is 2% higher than that of Retina LP alone.The detection accuracy rate of the license plate location model is 98.6%,and its detection speed is 39 FPS.2.In the section of the license plate character correction,it eliminates the influence of abnormal characters on license plate character recognition.First,we use the license plate fine localization algorithm Retina LP to detect the corner coordinates of the license plate.The method achieved 98.5% accuracy by testing 1000 tilted license plate images.In the link of the license plate character recognition,the conventional license plate character recognition algorithm needs the link of cutting character and weak robustness.This paper put forward the new license plate character recognition algorithm LR-VGG based on RCNN;it is based on the VGG13 structure,the activation function is replaced by tailoring and optimizing the network structure.Secondly,the feature information is learned using a two-layer bidirectional GRU module and the CTCoss loss function trains the LR-VGG model.The final experiment shows that the model converges quickly after using the CTCloss loss function,and the accuracy of license plate character recognition is 95.6%;when the license plate correction link is added,the accuracy reaches99.2%.3.In this paper,the modified Faceboxes rough license plate location algorithm,the modified Retina LP license plate character correction algorithm and the modified LR-VGG license plate character correction algorithm are combined to construct 2000 test data sets containing frontal,low-light and foggy images.Finally,in order to verify the rationality and effectiveness of our model in the paper,an ablation experiment was performed,the test set was tested with other algorithm models and the comparison results were obtained.The results showed that the accuracy rate of license plate recognition in this paper was 98.2%,and the detection speed was 29 FPS,which can match the requirements of real-time license plate recognition. |
PDF Full Text Request