Handwritten Kanji recognition using combined complementary classifiers in a cascade arrangement

In classifier combination, the degree of recognition accuracy improvement depends not only on how classifiers are combined, but also on how much they complement each other. From this view point, a complementary classifier that is combined with an existent classifier should use features that make it easy to recognize patterns misrecognized by the existing classifier. As such features, I propose using common difference principal components and difference principal components in a complementary classifier. I also propose combining complementary classifiers in cascade. In each complementary classifier, the common difference principal components and the difference principal components are obtained to make it easy to recognize patterns misrecognized in the preceding combined classifier. These ideas were applied to handwritten Kanji recognition. Experiments showed that the recognition rate increased from 97.35% to 98.00% after a single complementary classifier was combined and to 98.23% after four were combined