An Efficient Method Of License Plate Location In Complex Scene

License Plate Location(LPL)plays a vital role in the license plate recognition system.Under special environment,such as the bayonet area of toll station,traffic flow monitoring system and so on,the locating accuracy in such environment is quite high(more than 97%).However,when facing complex natural scene,the accuracy is still low.Under such complex environment,the license plates have different angles and illuminations,and there also exist many suspected license plates.The traditional methods often lead to an increment of error detections.In this paper,we propose a method which combines deep learning,machine learning and traditional vision methods to locate license plates in complex natural scenes.The presented method is divided into three steps:the first step is vehicle location.Obviously,the license plate is part of the vehicle.Once the car regions are found,it will greatly decrease those error detections in regions where there are no vehicles.In the second step,we first use crawler to crawl a certain amount of license plates as our initial samples,then we will utilize these initial samples to generate more samples by adding noises,projection transformation and etc.Based on these samples,we trained our cascade classifier using AdaBoost with LBP features and decision tree with traditional computer vision features separately.In the last step,we first use cascade classifier to get candidate license plate areas,and then use decision tree to classify these candidate areas,after these two procedures,we get our final license plate regions.The experimental result shows that the present method can greatly reduce the number of error detection areas.Meanwhile,it also guarantees the accuracy of license plate location.