A Bayesian Approach for Licence Plate Recognition Developed on a Realistic Simulation Environment

Abstract

Template matching is one of the most common methods for license plate recognition. This method discards prior probabilities of license plate codes. The posterior code class probabilities constructed by including the prior probability information are expected to improve the recognition performance. The probability information that needs to be included requires extensive training data, which is quite costly to obtain. In order to generate these training images a license plate image simulator is developed with a realistic noise model. Simulated license plate images are then used to test a Bayesian decision theory based recognition procedure. Test results indicate that, with the inclusion of prior information, significant recognition gain is obtained with respect to standard template matching method at high noise levels.