The Research And Implementation On License Plate Character Recognition System

License Plate Character Recognition is the core of License Plate recognition and a critical technology step in Intelligent Traffic System. And now it is one of the important research topics in this field.The blue bottom and white word license plates got by making location are the processing samples in this paper. The paper has a deep research in the important composing techniques of character recognition. These are the character segment,the character feature extraction and the design of classifier machine. The main work is as follows:1. The part of character segment: Based on the license plate pre-processing, such as: greyscale, using Otsu method to change greyscale image into two-value image and removing the license plate frames and up-low rivets, the paper presents an improved horizontal projection based on the prior knowledge of license. The experiment shows this method is effective.2. The part of character feature extraction and optimization: After the processing of smoothing and normalization, a comprehensive method is proposed based on the analyzation of the usual methods of feature extraction. The method is extracting robust grid and robust contour character + making feature reduction used Principal Components Analysis ( PCA).3. The part of character recognition based on BP network: In this paper an improved BP network imposing momentum gene,a self-adaptive learning rate is proposed after the research of how to design the network structure,the training parameters and the analyzation of the advantages,disadvantages while running the BP network; Based on the characters of the license plate images, we have designed four kinds of classifier machines which are used to recognize the Chinese,English,Number,English and Number. Then the correct recognition rate based on the features got by extracting robust grid and robust contour character has a compare with the very rate based on the optimization features reduced by using PCA. The experiment result shows that using the optimization features the BP network running speed and the correct rate has been improved.4. The exploration of Chinese character recognition based on LSSVM: Considering the number of Chinese samples,the complex structures of Chinese and the advantage of SVM dealing with the less samples, we have a exploration of Chinese character recognition based on LSSVM. First, construct Multi-classifiers of LSSVM based on"one-to-all"and RBF kernel function. Two kinds of different feature dimensions are compared their correct recognition rates. The experiment results shoes that the Multi-classifiers of LSSVM can get higher recognition rate based on the features extracted by robust grid and robust contour character without using PCA.