The Research And Practice Of License Plate Recognition Algorithms

With the rapid development of domestic economy, the number of vehicles increases quickly in our country. It demands higher level of traffic management. Intelligent Transportation System (ITS) develops rapidly. It applies information technology into transportation system while improves greatly the efficiency of transportation and the level of management. The identification of each vehicle is an important content of ITS, while license plates are the chief symbols of vehicles. The technology of License Plate Recognition (LPR) has to adapt outdoor and all-weather working environment, and deal with all kind of plates in the actual scenes, which may be blurry, dirty and slant. At present there is not a system can be applied generally in the world. This shows the diversity of license plate and the complexity of the environment while LPR have to deal.On the digital images of vehicles in traffic scenes, this paper researches different phases of LPR algorithm. We design and implement a LPR system, which input vehicle images and output the text result of license number. The LPR algorithm divides into three phases: License Plate Locating, Characters segmentation and Characters Recognition.In License Plate Locating phase, this paper synthesizes gray processing and the usage of color information. First we detect the vertical edges of the whole image, and then we use line scan technology to locate plate regions based on characteristics of edge pixels in those regions. After that we calculate the color value of pixels in the plate regions, and get the color type of the plates. Based on the result we judge the validity of the plate regions. These are lots of plates which are skew in the images, so we have to correct them after locating.Characters segmentation means that a series of single character image are extracted from plate region images. Here we propose a method of projection to find the boundaries of characters. There are many problems we have to deal with, such as the disturbance of plate frame, vehicle body or image noise. We solve those by using the size information of characters and array rule, and get the character images which have accurate boundaries.On the character recognition phase, we extract the characteristics first. After the characters are normalized, we extract a part of structure characteristics and statistic characteristics as characteristic space. Then we design a multi-class geometrical classifier to decompose characteristic space into sub-space based on different classes. This classifier is effective when we use it to train or recognize characters.The result of experiment shows that the algorithms we use in this paper have high rate of plate recognition. And the executing speed is fast. The algorithms can be applied to actual projects. The system we implement in this paper is a valid LPR system.