An Improved Approach to Generating Realistic Kanji Character Images from On-Line Characters and its Benefit to Off-Line Recognition Performance

This paper proposes a method for generating realistic calligraphic Kanji character images from on-line data. The proposed method is an improvement of our former method presented in [1]. Our new method can cope also with connected on-line strokes, i.e., stroke number variations, which were not correctly painted in our previous method. The new method decomposes strokes into three different parts (end part, bend part, connecting part), and paints each part according to a prototypical shape assigned to it from a calligraphic stroke shape library. Our generated calligraphic off-line images serve two purposes: First, they provide additional training samples for off-line recognition. Second, they allow application of off-line methods in on-line recognition. This paper also presents some experiments for these purposes.