In the processing of Korean character recognition, scanned binary character image is thinned to extract strokes to be compared with character models in order to determine the class of the character. But noises, i.e., deletion or addition of short stroke elements in thinned binary character images, often cause misrecognition.
In this thesis, we propose a character recognition algorithm with new internal representation of phonemes. An input character image is represented as a sequence of connected components from preprocessing. A phoneme is extracted by matching a connected component with phoneme models and the extracted phoneme is used to construct a syllable. This process is repeated until we arrive at a complete syllable. But when it reaches a deadend, it backtracks onto previous matching, and another phoneme model is tried. During that process, phonemes split due to noises may be relinked through an expectation-based processing. Experimental results revealed this technique has enhanced robustness in the recognition of the Korean characters.