The Research Of License Plate Recognition System And Its Learning Mechanism

The technique of License Plate Recognition(LPR) becomes more and more mature because of the rapid development of Intelligent Transport System(ITS) . Currently more and more research institutes and commercial organizations work on the research of LPR , and they have made much process but they also have their own limitations. The research institutes mainly focus on testing algorithms or publishing papers, and the systems developed by them are rarely used in real project, and the stability and adaptability are not so good. The system developed by the commercial organization has great function and it can be applied in all kinds of environments very well, but its algorithms can't be updated quickly and its system performance can't speed up fastly. The algorithms are commercially protected, which doesn't benefit the communication and improvement of LPR.Our lab also has done many related algorithms research for nearly 10 years but a perfect system hasn't formed up to now. So we hope to develop a complete license plate recognition system which has a clear function modules and some feedback and learning ability. Our works mainly focus on the following aspects:A complete architecture of license plate recognition system are proposed,which include the learning mechanism, and the function of modules are divided. A character segmentation method combining the projection and instinct characteristics of license plate is proposed during the license plate location period. The rough license plate is located firstly, then the prior knowledge of license plate characters is used to judge whether the characters are located correctly. By adjusting the characters sequence,the correct characters will be located at last. The experimental result shows that this method has a good effect on character segmentation. According to the distribution of the characters, a thinning module method are proposed after comparing and testing many kinds of characteristics. The results shows that these characteristics have an effective distinguishing capability and they also does well to the familiar characters. The SVM classifier is used to recognize the characters. SVM classifier also has a good effect on the small sample problem. Those characters which have been recognized to wrong characters will be returned to the corresponding period after the learning mechanism is included. The efficiency of the system will be greatly improved after training and studying again and again under the supervised learning way.