License Plate Recognition in Diversified Situations Using Robust L-GEM-Based RBFNN

Abstract

The most critical step in license plate recognition tasks is the identification of individual character image from the license plate image segments. Conventional methods of recognizing a character including Support Vector Machine (SVM) and neural network require the training using many license plate images. However, the amount of training data is limited and there are many unseen situations, where the generalization capability of a trained classifier is usually limited. If the license plate image distortion is serious due to either weather conditions or technical reasons of photographing, accuracy of these methods will be greatly reduced. Therefore a robust license plate recognition method is proposed using a Radial Basis Function Neural Network (RBFNN) trained via a minimization of the localized generalization error model (L-GEM). The L-GEM provides the upper bound of the generalization capability of an RBFNN with respect to a given training data set. Therefore, the trained RBFNN yields a better generalization capability and a higher recognition rate for new unseen samples. Experimental results show that RBFNNs trained by minimizing the L-GEM always yield the highest accuracy in diversified situations, such as rainy and snowy conditions.