Research On Recognizing Vehicle License Plate Character Under Natural Environment

Vehicle License Plate Recognition (VLPR) system has been widely used, which is important to the Intelligent Transportation Management System (ITMS). Vehicle license plate character recognition, as one of the most important steps in the VLPR system, is also a hotspot in recent research. It combines several kinds of technology including image processing, pattern recognition and machine learning.Under natural environment, the plate character can be faint or transfigured. As a result, we bring up several algorithms to process the noise in the image, including single point noise processing algorithm, noise processing methods based on connected area, second-moment and edge scanning, which are efficient in getting rid of the noise. To describe the feature of the image more efficiently, we compare some methods of feature extraction and adopt the gridding barycenter feature as the feature of the character image.Because of the lack of training samples, traditional methods based on experiential risk minimization can not play well in recognizing the characters. Statistical Learning Theory focuses on the rule of machine learning with small sample sets. Support vector machine is a new generated machine learning technique based on VC dimension and structural risk minimization. So we use SVM to recognize the plate characters according to the reasons above. To design the classifier based on SVM, we put forward the standard to estimate the performance of parameters: predicting accuracy and number of support vectors. Then we propose the method of second gridding searching to get best kernel parameters. Using this method, we get best parameters of poly kernel function, RBF kernel function and Sigmoid kernel function. We get best kernel and parameters from them. Meanwhile, we present the method of designing multi-class classifier. To solve the problem of some images'poor quality, we give the standard to estimate the performance of utterance rejection method: utterance rejection accuracy and recall rate. And we bring forward the utterance rejection method based on voting threshold and analyze the performance of our algorithm.