Research On Bangla Handwritten Numeral Recognition Using Decision Tree And AdaBoost

Based on the project of Bangla Post Automatic Letter Sorting Machine manufactured by China Postal Corporation, in this dissertation, we implements a Bangla handwritten numeral recognition system using decision tree and AdaBoost algorithm. This system is aimed at obtaining the high reliability recognition performance for Bangla postal numerals. The system consists of image preprocessing, feature extraction, feature discretization, two-stage classifier based on AdaBoost and decision tree. Experimental results show that the method is of high reliability and strong robustness in recognizing Bangla handwritten numerals and meets the practical requirements.The main content and structure of this dissertation are as follows:1. Image preprocessing on Bangla handwritten numeral images is implemented, including image smoothing, binarization, stroke width normalization, character size normalization, etc.2. According to the special structure of Bangla handwritten numerals, open-loop and closed-loop features are employed. These features are able to describe each Bangla numeral category and distinguish one from others, which reduce the dimensions effectively and greatly improve the efficiency of the classifier.3. Statistical features, i.e. directional features are extracted as a complementary of the above structural features. This kind of feature is of good robustness on noisy images.4. Probability estimation decision tree with Laplace probability smoothing is adopted as the base classifier of the system. It forecasts the numeric categories, and provides the information of reliability at the same time.5. AdaBoost algorithm is applied to Bangla numeral recognition. It upgrades the decision tree classifiers to a strong combination classifier, and improves the recognition performance significantly.6. A two-stage combination classifier is designed based on the above two complementary features. And the reliability of recognition results is controlled by two predefined thresholds. The characters rejected by the first-stage classifier are recognized by the second-stage continually. Therefore, the system achieves high reliability and recognition rate.