Research On Deep Learning Based Intelligent License Plate Recognition System Under Complex Scenes

In recent years,artificial intelligence technologies represented by deep learning have brought unprecedented changes to all industries.As an important part of "smart city" and"smart transportation",license plate detection and recognition play an increasingly important role.In the era of artificial intelligence,traditional license plate detection and recognition schemes based on manual feature extraction can no longer meet the needs for a more intelligent and faster lifestyle.Based on the above consideration,this thesis studies the license plate detection and recognition algorithms based on deep learning.The intelligent license plate recognition system suitable for complex natural scenes is studied and implemented.The research contents of this thesis mainly include the license plate detection algorithm,the license plate character recognition algorithm and the design of license plate intelligent recognition system.Inspired by the single shot multibox detector(SSD)algorithm,this thesis proposes an improved SSD for the detection of Chinese license plates.The improved SSD algorithm enhances the detection performance of the original SSD algorithm for small targets such as license plates in complex scenes by fusing feature information of different scales,deepening the depth of the prediction layer,and setting appropriate default boxes that match the size of the license plate.The experimental results show that the proposed algorithm achieves higher detection accuracy than some existing ones.For the license plate character recognition task,this thesis presents a sequence recognition algorithm of Chinese license plates.The algorithm effectively improves the recognition accuracy of the algorithm in complex scenes through the modification of spatial transformer network(STN)module and the feature extraction of improved convolutional neural network(CNN)module.The features of different convolutional layers are integrated as input to a bi-directional recurrent neural network(BRNN),where the character segmentation is not needed.Finally,the recognition is accomplished by the BRNN and connectionist temporal classification(CTC).Due to the lack of adequate Chinese license plates,an effective training method is presented.Experiments prove the rationality and effectiveness of the recognition algorithm and training method.Based on the improved SSD algorithm and license plate recognition algorithm,this thesis proposes two schemes for designing license plate intelligent recognition system.One is the one-stage license plate intelligent recognition system and another is the two-stage license plate intelligent recognition system.The two-stage system performs license plate detection and license plate character recognition as two separate modules,while the one-stage system combines the two modules into one network.The advantages of the two systems are summarized through the comparative analysis of specific experiments.The two-stage system shows the higher recognition accuracy and reliability,while the one-stage system has advantages in terms of speed,easy-deployment and portability.