Generating Chinese Character Skeletons with Parallel Random Recurrent Networks

Authors:
Publisher website
Google Scholar
Abstract
Chinese character skeletons are key topological description of Chinese character glyphs which provided essential information about structures of the Chinese characters. The Chinese character skeleton can be regarded as sequential points generated using deep neural networks, such as recurrent neural networks (RNN) while time intervals between sequential points which included dynamic information of handwriting process of Chinese characters could not be presented elegantly using RNN. Deep neural networks with long sequence outputs face problems such as gradient vanishing or explosion. Chinese characters had longer sequential points than those of English characters. Here, we generated the Chinese character skeletons using multiple parallel random recurrent networks (RRN) which provided rich dynamic presentation of sequential points. The generation took place at two levels, the stroke level which included a sequence of points, and the character level which included a sequence of strokes. Our method not only generated temporal information of the skeletons but also retained more skeleton feature points with only small-scale training data which made the generated Chinese character skeletons had richer writing detail features than existing methods. Our method avoided the problems of gradient vanishing or explosion, and could be used to quickly generate large-scale and high-quality Chinese character skeletons.

This publication has 8 references indexed in Scilit: