License Plate Number Recognition System Based On The Handheld Terminal

With the rapid development of economy and society,the number of motor vehicles shows explosive growth,which led to the intelligent and automatic management of vehicles more and more important,so license plate recognition system in vehicle management plays a vital role,and becomes an important research direction of intelligent traffic.At the same time,the breakthrough development of handheld intelligent terminals regarding the mobile phone as the representative,promote the development of society and provide a new support platform for the application of license plate recognition system,because of its portability and information exchange capacity.The license plate recognition system based on the handheld terminal is built on the advantages of both the handheld intelligent terminal and license plate recognition system,which uses the handheld terminal as the carrier of image acquisition and information interactive display and adopts the deep neural networks in the field of computer vision to realize the image recognition,to build up the new license plate recognition system.In this paper,the license plate recognition system based on hand-held terminal is divided into three modules: license plate location,character segmentation and character recognition.Considering that the images collected by the handheld terminal are subject to some problems,image distortion correction and illumination normalization are added.The main works of this paper are as follows:1.We established a new license plate location and segmentation model based on fully convolution deep neural networks.The license plate location and segmentation model uses the fully convolution neural networks,which extends the semantic segmentation function of the fully convolution networks,to realize the identification and segmentation of the specific license plate target.And uses the handheld terminal to collect vehicle images,to build a test data set with more than 4000 images.In addition,we use the Markov random field to combine the localization method based on the image color feature and the localization method based on the edge information of the image texture,to improve the traditional location and segmentation method as the comparison of license plate location segmentation method based the deep neural networks.The comparison experiments show that the location and segmentation method based on fully convolution deep neural network has better location effects and operational advantages.2.We propose a character segmentation method based on the priori waveform characteristics of license plate character.This method extracts the character of the projection curve as the basis for the license plate character segmentation,which is a more flexible license plate character segmentation means with the effective character segmentation.In addition,the character segmentation result is compared with the a priori waveform characteristics of the license plate character,which sets the feedback mechanism to reduce the wrong character segmentation.3.We propose a method for correcting the distortion of the license plate based on the perspective projection transformation,and the illumination normalization method based on homomorphic filtering.This method mainly focuses on the background interference,illumination change,image angle scale transformation and other problems of the license plate,to improve the robustness of the license plate image collected by the handheld terminal,which is beneficial to the correct segmentation and recognition of the license plate characters.4.We build a license plate character recognition network based on deep neural networks.The network combines the TensorFlow platform,Google’s advanced learning platform,to build a license plate recognition system.The network combines the TensorFlow,a deep learning platform introduced by Google,using a convolutional neural network structure,and the license plate character library for training and testing.Compared with the traditional license plate character recognition,the deep neural character recognition network has a higher recognition rate.