A hybrid method for unconstrained handwritten numeral recognition by combining structural and neural "gas" classifiers

This paper presents a hybrid method for handwritten numeral recognition that combines two compensatory recognition algorithms by analysing their performance for several aspects. The skeleton-based structural recognition algorithm employed in this method is robust under distortion but sensitive to noise and flaws. On the other hand, the neural network classifier, which uses scaled binary images as features and the neural "gas" model for classification, is relatively immune to noise and flaws but sensitive to distortion. The different performances of the two algorithms for broken, connected or slanted numerals, and the measurement-level decision provided by the neural network are detected and combined with different strategies to develop matching rules for each recognition method. Five combination methods based on performance analysis are developed to meet different requirements. As the two algorithms have fairly compensatory properties, the proposed method improves the recognition rate and reliability by exploiting the advantages and avoiding the weaknesses of each classifier. The experimental results from a large set of data show the efficiency and robustness of the proposed method.