Chinese Character Skeleton Extraction Based on Linear Span Network

Abstract

Chinese character skeletons are key representations which abstract structures of Chinese characters into skeletons. Extraction the skeleton from images of Chinese characters is a fundamental task in field of Chinese character process. Existing image skeleton thinning algorithms for extracting Chinese character skeletons produced bifurcations which could not accurately reflect topological features and writing trajectories of Chinese characters. Existing image skeleton extraction methods based on deep neural networks also had difficulty in obtaining accurate single-pixel Chinese character skeletons. Here we proposed a hybrid Chinese character skeleton extraction method to resolve the above problems. Firstly, a dynamic skeleton extraction algorithm was proposed to automatically obtain a dataset using a customized digitalization platform of Chinese handwriting process. Secondly, an improved linear span network was trained using the dataset to extract the Chinese character skeletons. Thirdly, the extracted skeletons were refined using an image-thinning algorithm to obtain single-pixel Chinese character skeletons. Results showed that our method not only preserved the topological structures and writing trajectories of Chinese characters but also maintained connectivity at intersections of strokes which avoided generating burrs and bifurcations.