Research On Algorithms Of Vehicle License Plate Recognition Based On Machine Learning

With the development of the city and the extension of roads,it is impossible for the city management to spend as much money as it used to.The smart cities,intelligent transportation and other related concepts are ready to emerge.The vehicle license plate recognition system has played an important role in routine situations such as access control system and road supervision.However,in unconventional scenarios such as honking capture and vehicle borne video recorder,the precision of vehicle license plate recognition still needs to be improved.We study related algorithms on vehicle license plate recognition system,and improved system's performance in unconventional scenarios by using machine learning algorithms.For the convenience of description and research,we decompose the license plate recognition system into three modules: license plate detection,character segmentation and character recognition.1.An improvement method to the Fast Feature Pyramid Cascade(FFPC)for vehicle license plate detection module.The role of this module is to detect the presence and location of license plates.In order to solve the problem of poor self-adaptability and slow detection speed of existing license plate detection algorithms,this paper improved the FFPC detector algorithm and significantly improved the license plate detection speed and self-adaptability.On this basis,we propose a vehicle license plate verification network based on deep learning,which improves the accuracy of detection with losing little the recall rate.In addition,in order to further improve the quality of the detection results,a precise positioning algorithm is proposed,and it also provides a new idea for image pixel segmentation.2.A multi-frequency global matching algorithm for character segmentation.The role of this module is to separate seven characters from the detected license plate image,and the module is divided into two parts: tilt correction and character segmentation.In the tilt correction section,we study a projection correction method,and create a better tilt metrics,and result in improving the adaptability to large-angle license plates.In the character segmentation section,a multi-frequency global matching algorithm is proposed for the problem of inaccurate segmentation.And experiments show that the proposed algorithm obtains significant results,especially in unconventional scenes.3.Character recognition module based on a modified generative adversarial network.For the problem of low character recognition rate,we propose a novel method for extending the character dataset of vehicle license plate,which could provide more samples for the later training process.We use secondary classification support vector machines to recognize characters,and find that the extended samples help improve the recognition accuracy.In addition,the data set expansion method proposed in this thesis does not require manual labeling,which greatly reduces the research cost,and can be flexibly modified to generate the character samples according to different requirement,and has the promotion significanceRelevant experiments in this paper run in the Intel Core i5 processor.In the case of special circumstances,it will be particularly stated in the relevant parts.Because there is no public license plate test data set,the experiment results are all based on our own test data set.The optimized algorithm will be compared with those basic algorithms.A large number of experiments have proved that the optimization has made the system more robust.