A Font Style Learning and Transferring Method Based on Strokes and Structure of Chinese Characters

Abstract

The existing method of contour-based font description is difficult to meet the personalized need for various style font generations because of the large size of Chinese character set. In this paper, we propose a novel glyph description method which treats the Chinese character as a constitution of the stable part called "structure" and the mutable part called "style". The structures of all characters are clustered by an improved K-Medoids method to guide the following generation of sample set which covers all kinds of style information of the whole character set. The result of cluster procedure indicates that radicals are bottlenecks for the reduction of sample set due to the low repetition rate in all characters. To address this problem, we present the radicals as a set of stroke-to-stroke layout structures, and render them from these substructures available in the sample set. Experiment results shows that the substitution enables us to learn the style information from a small set of sample characters (less than 10% of total amount) and generate the rest with the similar writing style.