License Plate Character Segmentation Using Key Character Location and Projection Analysis

Character segmentation forms a link between license plate detection and character recognition. This paper presents a two-stage character segmentation framework. Firstly, a cascade classifier is trained by AdaBoost algorithm to locate key character which represents an administrative area. Then, based on key character location information, the rest of characters can be predicted or estimated. For the neighbor character of key character, it can be predicted by connected components analysis. For other characters that are far from the key character, a vertical projection strategy without hyphens is proposed to segment them. This strategy is promising to address the issue of touching characters. To illustrate the effectiveness of character segmentation, broad learning system is adopted to classify segmented characters. Experimental results performed on Macau license plates demonstrate the proposed method's effectiveness in comparison with some state-of-the-art approaches. It is expected to apply the designed techniques to license plate character segmentation of other regions or countries that have similar situations with Macau's.