License Plate Detection Based On Deep Learning

Recent years,more and more people have thrown themselves into the research on theory and application of ITS(Intelligent Transportation System)in the world.License plate recognition technology become the core technology of intelligent transportation system,in order to achieve the key aspects of intelligent has been the rapid development.It is required artificial intelligence,image processing,computer vision,pattern recognition,deep learning and other related technologies integrated application.In the entire license plate recognition system,license plate detection is one of the more critical step.The traditional license plate detection algorithms have some limitations in some scenes.Especially in low light conditions,when the license plate texture or color similar to the background,the traditional license plate detection algorithm will have a false test.This paper is from this point of view,proposed improvements or new methods make use of machine learning and deep learning related technologies.The main contents of this paper include:1,discuss the traditional license plate detection algorithms and analyzes their advantages and disadvantages.The traditional texture algorithm is sensitive to noise and when encountered other similar to the license plate texture characteristics of the interference,there will be the phenomenon of reduced license plate detection rate.The traditional color algorithm when the image exists with the color similar to the license plate,the license plate detection accuracy will decline.According to the previous experimental research and related machine learning methods,an algorithm based on color differential model combined with Adaboost for license plate detection is proposed.It combines with the advantages of color algorithm,texture algorithm and original Adaboost algorithm.The characteristics of the license plate are carried out Adaboost training,to solve the traditional license plate detection algorithm problems.2,this paper is proposed the license plate detection algorithm based on fast RCNN deeply.First,select the license plate candidates by selective search.Then,fast R-CNN directly on the entire license plate image convolution to obtain feature maps.Secondly,the license plate candidates are mapped to the feature maps to obtain the corresponding feature of each candidate box.Since the size of the original image is not limited,the ROI pooling layer is added to normalize the feature dimension.Finally,the network directly outputs the boundingboxes location(ie,the license plate detection area).In the process of the network model training,the networks learn almost all the key features including the texture,color and difference of the license plate.In addition,the multi-task loss learning method improves the accuracy of the license plate detection.3,the experimental part is mainly for two traditional license plate detection algorithm and two algorithms proposed in the paper to test and compare.First,we use the algorithm based on RGB color space to test the license plate image.It is found that it is more sensitive to the change of illumination.Secondly,we use the algorithm based on texture feature The license plate image to achieve,through experiments found that similar to the license plate texture interference will appear the license plate error detection.Finally,the algorithm based on the color differential model combined with Adaboost algorithm and FRCNN algorithm are used to test the license plate.The experiment proves that both of these algorithms can solve the problems existing in the traditional license plate detection algorithm under weak light,and can accurately detect the license plate.