Research On Key Techniques Of Low-Quality License Plate Recognition

With the development of information technology and intelligence technology, the informatization and intelligentization of traffic management is the direction. License Plate Recognition system (LPR) is the core of Intelligent Traffic System (ITS). It is very important in modern traffic management systems.In this paper the three main parts of LPR systems are discussed comprehensively including license plate location, character segmentation and character recognition. A deep research is made especially about some important and key technologies in low quality license plate recognition.In the license plate location stage, a prior knowledge-based method is proposed. Research is made mainly aiming at the location of low quality license plates under various kinds of poor conditions. The prior knowledge such as the size of the plate, the character width, the character height, the distance between characters and the stroke width is fully made use of. Firstly, 2-dimension wavelet transform (DWT2) is used to extract the edges of the strokes. Secondly, morphological operation and geometrical features are used to locate roughly and get the candidate regions. Finally the fine plate region is located through the prior knowledge of the distribution feature of strokes and characters.In the character segmentation stage, a prior knowledge-based algorithm which is robust to poor images is presented. The prior knowledge which can be used in the segmentation is summed up. Radon transform is used to perform incline correction. Raw segmentation is carried through the vertical projection of the binary image of the plate. The fine segmentation result is attained by the full use of the prior knowledge.In the character recognition stage, a method which is based on SVM is proposed. The multi-class classification problem is transformed into the binary classification problem through the "One Against All" mode. Different classifiers are set up for different kinds of characters. The method supports small training sets and works better in terms of training and recognition speed also recognition rate.Experimental results show that the method proposed in this paper is robust and practical, it can achieve good location, segmentation and recognition results even when the quality of the license plate is poor.