Conventional character recognition systems are usually employed either top-down (model-driven) or bottom-up(data-driven) control strategy. Since either strategy has its own merit, it is desirable to use a mixed strategy in which bottom-up and top-down approaches are used intermixedly to achieve a high degree of flexibility and efficiency.
In this thesis, we propose an algorithm for the recognition of Korean characters with a mixed control strategy. In the bottom-up stage, each character pattern is classified into one of the six fundamental Korean character types to reduce search space in classifing character patterns. In the top-down stage, the location and shape of each phoneme in the character is predicted and confirmed to recognize component phonemes. Feedback loops are provided to correct misclassifications that occurred at an earlier stage.
In a test case with the most frequently sued 522 printed Korean characters in Myungjo font, it achieved 97.2% correct recognition rate in the determination and 82.6% correct recognition rate in character (syllable) recognition. With the feedback process, its recognition rate was increased by 2.9%.