Research On SVM Multi-Class Classification Key Technologies And Its Application On License Plate Characters Recognition

Intelligent Transportation System (ITS) plays an important role in public transport management, security and other issues. License Plate Recognition system (LPR) is the core of Intelligent Transportation System (ITS), so it is significant to study license plate recognition technology.Support Vector Machine (SVM) is one principal aspect among license plate character recognition methods. Traditional SVM is designed for binary classification problems. How to well apply it into multi-class classification problems is an important part of SVM’s research. In this paper, we mainly study the key technologies of the SVM classification, and apply them into license plate characters recognition.The main contributions of this paper are as follows:1) A feature extraction method of pixel density features on annular region is proposed. Three kinds of features of a license plate character are extracted, including pixel density features on annular region, local grey level and gross periphery features. These features are then combined into a feature vector. Then the feature vector acts as an input for a multi-class SVM classifier. Pixel density features on annular region reflect the distribution of character pixels in the internal regions and the outlying regions in a character image. Experiments show that the proposed multi-feature extraction method has more advantages than some existing algorithms.2) An improved error-correcting output coding multi-class classification method is proposed, and a quaternary coding multi-class classification is also proposed. Error-correcting output coding (ECOC) is a SVM-based multi-class classification method, but this method, when the number of categories is not 2" (n is an integer greater than zero), has unclassifiable region. In this paper, we eliminate the unclassifiable region through using several key second-class classifiers. When ECOC code length is too long, the generalization ability of ECOC multi-class classification will be poor.To solve this problem, a quaternary encoding multi-class classification method is given, in this paper, to condense the length of ECOC code and improve the generalization ability of the classifiers. 3) A character recognition algorithm, which is based on the fusion of quaternary coding classifier, error-correcting output coding and minimum distance classifiers, is given. This algorithm uses parallel multi-classifier fusion mode. When recognizing a character, the combination of the pixel density features on annular region and features of local grey level can be used as an input for a quaternary coding classifier, meanwhile gross periphery features for an error-correcting output coding classifier and the features of all pixels for a minimum distance classifier. Finally, we fuse three recognition results with improved voting method to gain the final recognition result. Experiments show that the proposed method, in contrast single classifier, has more advantages.In this paper, we mainly study the key technologies of SVM classification, feature extraction and character recognition in license plate character recognition, and apply of SVM classification and multi-classifier fusion into license plate characters recognition.