CHARACTER RECOGNITION FOR LICENSE PLATE RECOGNITION TRAFFIC CAMERA IN VIETNAM

Abstract

Optical Character Recognition (OCR) is an active research direction with many practical applications, including digital character classification for license plate recognition on traffic cameras. The OCR models usually deploy a single classifier for all the categories in the dataset. However, the classification difficulties among all the classes in the dataset are different, some characters are easier to be misclassified compared to the others. Due to this reason, the classification performances across the classes are not equal. In this paper, we deploy a 2-stage classifier in order to improve the classification accuracy for difficult classes. The first classifier is used to classify all the classes while the second one is used only for difficult classes, in order to refine the predictions made by the first classifier. The experiment results on two datasets SVHN and license plate characters demonstrate that the proposed method helps to enhance the classification accuracy of some difficult classes by 1.4%.