A Structure Aware GAN-Based for Ancient Chinese Calligraphy Style Transfer

Abstract: A written Chinese brush calligraphy is an artistic style. Calligraphers use brushes to write characters for their artistic expressions and creations. They exhibit various structures and attractive stroke shapes. Unfortunately, brush-written characters of ancient masterpieces are getting unclear or damaged if their papers or steles decay. Unlike the alphabet, each Chinese character has its own significance and completion. Besides, calligraphers are used to write non-standard character forms for some Chinese characters. The different character forms are due to handwriting habits. Therefore, we propose a novel mechanism that reduces the effect of non-standard character forms by normalizing the loss value of the generated Chinese calligraphy. A generated calligraphic characters strictly followed the features of the structures and stroke shapes that is challenging for the generation procedure.Recently, the development of convolutional neural networks (CNNs) has enabled to generate font automatically. Some attempts have been made to learn font generation and achieve talented results. Many researches with deep neural networks have generated whole sets of alphabetic languages. So far, some researchers provided strokes and radicals to assist Chinese-font generation. However, only a few studies have focused on ancient Chinese calligraphy generations. This paper proposes a structure aware consistency framework to provide structural correction for the generative model. According to ancient Chinese calligraphy copybook, we developed a method for multi-style calligraphy transformation. An inverse mapping network is used to automatically supervise the structural correctness of forward-generated Chinese calligraphy characters. Finally, the proposed method generated Chinese characters that ancient calligraphers had never written. In generative Chinese calligraphy, previous works provide productive results. However, their approaches usually rely on manual intervention such as radical decomposition. Any manual annotated data is labor intensive. Sometimes they fail in the stroke shapes or generate the incorrect structures for calligraphic characters. To address these problems, our proposed inverse mapping architecture to penalize incorrect structures of the generated characters. The inverse mapping architecture improves the applicability of our framework. We also provide an overlooked mechanism to decrease the adverse effect of non-regular character forms. The overlooked mechanism normalizes the L1 loss of a generated character by multiplying an overlooked weight. This mechanism brings an interesting aspect of Chinese calligraphy generation. The contributions of this study are as follows: (1) A proposed structural consistent architecture to overcome the most challenging aspect of generating calligraphic fonts. This architecture guides the structural correctness of the forward generative characters and significantly improves the applicability of our framework. (2) To reduce the influence of calligraphic characters which are different from standard character forms. We normalize the loss value of generated characters using an overlooked mechanism. Our one-to-one transfer model tends to overlook the calligraphic characters that are non-standard character forms during training.

Keywords: Chinese calligraphy, Style Transfer, Generated character, Regular Script, Deep Learning

DOI: 10.54941/ahfe1007026

Cite this paper:

​