Research On The Improvement Of Plate Character Recognition

License Plate Recognition system (LPR) is a appropriative computer vision system, which use vehicle license as its recognize target. The system use computer vision and pattern recognition and other technologies widely. And it's one of the important research topics in Intelligent Traffic System. During the recognize process, the character split and recognition are the most important part. It directly reflects the ultimate recognition accuracy of the entire system. Also, the efficiency of these processes must be very high. In the real world, split and recognize a plate becomes a hard problem, because the intensity, color, direction of the ambient light changes observably, license plates region is polluted or partial deformity of characters, the angle between camera position and the plate also changes. This paper improve the split and recognize algorithm in two part: image pre-processing, and a character recognize algorithm based on support vector machine.In the license plate image preprocessing stage, image enhancement, Projection texture analysis and connection region analysis are used together to improve the result of plates inclination correction and split algorithm. The algorithm process gray image first, using a Ostu's method to make clustering of the gray value, and scale the gray value of foreground and background respectively, to enlarge the contrast of the image, so we can get a better binary image later. Then, projection texture analysis and connection region analysis are applied to split the characters.In the character recognition stage, a scaled gray image is used instead of binary image to be the input data, because using binary image will lost lots of data, and this method can make better use of support vector machine. After that, this paper comes out an algorithm based on SVM called Multi-Dim Multi-Level Recognize Tree. When a vector is classified into multi-classes, A low dimension SVM is used first, classify the vector into a smaller set, then, a high dimension SVM is used to decide thich class it belong to. Compared with SVM set organized in one line or a binary tree, this method run much faster, and will lead to a better result in application.