Design of a recognition dictionary using artificially distorted characters

Abstract

AbstractOne of the serious problems in developing an optical character reader (OCR) is designing a dictionary of handprinted characters which requires a tremendous amount of character data for training. This paper considers training by artificially distorted characters instead of handprinted ones in the design of a dictionary. The feasibility and the usefulness of the scheme are examined. In the experiment, 26 alphabetical characters are employed. The characteristic loci proposed by Glucksman are used as the features, and the nearest‐neighbor method is used in the classification. The recognition result of test data indicated that the dictionary designed by artificially distorted characters has almost the same performance as the dictionary designed by handprinted characters. The correct recognition rates are 98.6 percent for the former and 98.9 percent for the latter. By the training where a part of the handprinted characters are replaced by the artificially distorted characters, a performance exceeding that of the training using only handprinted characters is realized, with the correct recognition rate of 99.2 percent. Thus, it is verified that the proposed method is a useful substitute to the method by collecting handprinted characters. It is shown also that by observing the distorted characters in the feature space, the stability of the features against distortion can be evaluated.