Design And Implementation Of License Plate Recognition System Based On Transfer Learning

License plate recognition system is an important part of intelligent transportation system.In the intelligent transportation system,the license plate recognition is to use the uniqueness of the license plate to identify and confirm the vehicle,so as to ensure the feasibility and convenience of the system.At present,the license plate recognition system is mostly used in the specific Angle of the shooting scene,such as the community entrance guard,traffic,security,highway toll station.However,in the natural scene,due to the influence of Angle distortion,different lighting conditions,shielding,different pixel ratio,license plate deformation,complex background and other factors,the license plate recognition system cannot accurately locate and identify the complete license plate information.In view of this situation,this study focuses on license plate recognition in natural scenes.The main contributions and research contents of this paper are as follows:A license plate recognition system based on deep learning is proposed.The traditional license plate recognition algorithm is no longer applicable to the multi-interference factors in the natural scene.The license plate recognition system based on deep learning can improve the robustness of the system in natural scenes by extracting and learning the characteristics of license plates in various complex scenes.This study consists of three modules:license plate location,segmentation and recognition.Compared with the traditional license plate recognition system,the proposed scheme does not require image preprocessing and license plate character segmentation,which not only improves the real-time performance,but also avoids the loss of accuracy caused by character segmentation.A license plate location and recognition algorithm based on transfer learning is proposed.The license plate location and recognition modules are implemented by SSD target detection and Xception image classification algorithm respectively.Considering the limited training samples and a large number of labeling costs,the transfer learning mechanism is introduced to improve the module optimization.Experimental results show that the proposed algorithm can effectively solve the cold start problem and improve the system generalization ability.On this basis,the character recognition model is used as the feature extractor of license plate background color recognition algorithm to realize double recognition of license plate characters and colors.Finally,the license plate location,cutting and recognition steps are integrated to build the server-side license plate recognition system,and the overall performance of the system is tested on the 5000 test sets.The test results show that the accuracy of the system is 98.34%,and the average recognition time of each image is 88ms,which better meets the needs of the license plate recognition system in the actual scene.