The Research On Techniques Of Carlicense Plate Recognition Based On BP Neural Network

With the rapid development of economy, the License Plate Recognition(LPR) of theIntelligent Transport System(ITS) has entered higher development stage. It has been widelyapplied in various occasions related to vehicle management, and occupies a place in thefield of traffic control system. The technology of digital image processing, BP neuralnetwork has been used in our research work. Same modules related to image preprocessing,license locating, character segmentation and character recognition in license platerecognition system have been studied. The experiments have been done by MATLABsimulation software.Firstly, considering the serious air pollution and haze weather in the city, somepreprocessing operation have been done to the collected foggy images including the imagedefogging, grey level transformation and image filtering. Emphasis is placed on the imagedefogging in this part. The algorithm of image defogging based on the prior dark channelhas been studied systematically. A modified algorithm has been proposed. The core idea ofthe modified algorithm is the improvement of the transmission rate. The experiment resultsindicate that the improved algorithm is better than the original on the image defogging,especially in the regions of the white and bright. The original characteristics of the colors inthese regions can be restored.Secondly, the shortcoming and deficiencies of the common method in the phase oflicense plate locating and segmentation have been explained in the corresponding modules.The horizon scans and vertical projection are combined to locate the license plate. The localvertical projection method has been applied to segment the characters by the aid of priorinformation of the license plate area. And the individual characters have been sent into therecognition module for the character recognition.Finally, the related theories and algorithm process of the BP neural network have beenintroduced. In consideration of the particularity of the license plate in our country, thecorresponding neural networks including Chinese network, letter network andletter-number network have been designed. The selections of the parameters in eachnetwork have been studied. And the three networks also have been trained. The experimentinsults show that the designed networks not only maintain the nonlinear mapping, self-learning and adaptive ability of normal BP neural network, and also enhance thegeneralization and fault tolerance ability. The recognition accuracy of the designed networkhas been improved.