Research On License Plate Location And Recognition Technology In Surveillance Video

License Plate Recognition(LPR)is one of the key issue in "Smart City"and intelligent transportation system.Although in recent years,the technology of LPR has made significant progress with the development of computer hardware and the evolution of deep learning,there is still much room for improvement in the problem of license plate recognition in surveillance video,such as the lack of training data,the complexity of envirnmental factors and the motion blur caused by high-speed motion.This paper focuses on the research of license plate recognition technology in surveillance video,which mainly contains the following work contents:Firstly,a labeling system based on EasyPR is designed for plate data,which can easily label the position and characters of the license plate in an image.This paper also designs a data generation method which can generate license plate data in various types.The data generation method not only guarantees that there is sufficient data for training,but also ensurs the diversity and blacnce of training data.With the training data generated,the accuracy of the license plate recognition network can be fully improved.Secondly,this paper analyzes the structure and processing of ConvNet-RNN,and the network is improved to achieve a segmentation-free recognition network for license plate.The proposed network AC-RNN has a significant improvement on the accuracy of license plate recognition compared with ConvNet-RNN.The main improvements of AC-RNN over ConvNet-RNN include the enhancement of semantic correlation between license plate characters,the application of attention mechanism,and the CTC Decoder optimized for license plate recognition.And then,a multi-task network is proposed for segmentation-free recognition for license plate recognition in this paper.This new method takes the character recognition of a position on the plate image as a sub-task,and learns all sub-tasks at a time to recognize all characters directly.The experimental results show that the multi-task plate recognition network proposed in this paper performs better than commercial license plate recognition algorithm.Next,based on the multi-task network for plate recognition,this paper proposes a phased training method to integrate all types of plate recognition processes together.The training method can make full use of the plate information that has been trained in the model,and guarantee the model's nrecognition ability for new plates in the absence of real training data.Finally,this paper designs a framework for plate recognition in surveillance video.First of all,SSD models and KCF algorithm are used to generate a sequence of plate images,and then all plate images are identified by the multi-task recognition model.At last,a series of integration strategies are adopted to generate the result of the plate sequence.The integration of plate image sequence further improves the accuracy of recognition,with satisfactory results in the realistic scenes.