中文字不同於其他主要語言，如：英文的26個字母、西班牙文的27個字母，正體中文僅僅常用及次常用字就超過一萬三千個，而每個字都是以部件為零件，依照不同的組字規則排列組合而成，因此如何克服大量的字數以及更有效率的設計字型是中文字型長久以來的問題。本論文將以風格轉換為概念，透過深度學習的方式解決這項問題。相較於以往字型生成的方法，在特徵提取部分需要較多的前處理且無法同時適用於印刷體及手寫體。我們提出一個基於卷積神經網路的中文字型轉換架構，該架構將直接以圖片格式作為輸入，透過以少量挑選的字訓練而成的類神經網路對輸入字型進行特徵的提取與轉換，並輸出為語義內容相同但字型風格相異的字。實驗評估結果顯示，本論文所提出的中文字型風格轉換架構，能有效轉換出近似於原始字型的中文字。

Traditional Chinese is significantly different from other major languages, as compared to western countries use Latin alphabet letters system, Chinese contains over 13,000 common and sub-commonly used characters. Each of these Chinese characters is composed of multiple parts, which are arranged and combined according to various orthographies. Therefore, how to handle these enormous character counts and how to design of more efficient fonts are longstanding issues in the creation of Chinese fonts. In this study, we seek to address these issues using deep learning based on the concept of style transfer. Previous font generating methods generally require multiple pre-processing steps for feature extraction, and these cannot be simultaneously adapted to typographic and handwritten characters. In this study, we propose a neural network architecture for Chinese fonts conversion, which directly uses images as inputs. In this architecture, neural networks are trained using a small number of selected characters, which are then used to extract and convert the features of an input font. The outputs of the architecture are characters with identical content albeit different font styles. We apply our method to ten different font style including typographic and handwriting and demonstrate it capable of converting Chinese-character fonts into fonts that are similar to the original fonts.

[1]Y. Jing, Y. Yang, Z. Feng, and J. Ye, “Neural Style Transfer: A Review,” arXiv:1705.04058, 2017.
[2]L. A. Gatys, A. S. Ecker and M. Bethge, “Image Style Transfer Using Convolutional Neural Networks,” In 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, 2016, pp. 2414-2423.
[3]Y. Taigman, A. Polyak, and I. Wolf, “Unsupervised cross domain image generation,” arXiv:1611.02200, 2016.
[4]S. Baluja, “Learning Typographic Style,” arXiv: 1603.04000, 2016.
[5]Master Ideographs Seeker for CNS 11643 Chinese Standard Interchange Code [Online]. Available: http://www.cns11643.gov.tw. [Accessed: June- 2017].
[6]CJK Strokes [Online]. Available:
http://www.unicode.org/charts/PDF/U31C0.pdf. [Accessed: June- 2017].
[7]Z. Lian, B. Zhao, and J. Xiao, “Automatic generation of large-scale handwriting fonts via style learning,” In SIGGRAPH ASIA 2016 Technical Briefs, 2016, No. 12.
[8]S. Xu, T. Jin, H. Jiang, and Lau FCM, “Automatic generation of personal Chinese handwriting by capturing the characteristics of personal handwriting,” In Conference on Innovative Applications of Artificial Intelligence, July 14-16, 2009, Pasadena, California, USA.
[9]B. Zhou, W. Wang, and Z. Chen, “Easy generation of personal Chinese handwritten fonts,” In 2011 IEEE International Conference on Multimedia and Expo, 2011.
[10]H.-C. Chen, L.-Y. Chang, Y.-S. Chiou, Y.-T. Sung and K.-E. Chang, “Chinese Orthographic Database and Its Application in Teaching Chinese Characters,” unpublished.
[11]D. Shi, R. I. Damper, and S. R. Gunn, “Offline handwritten Chinese character recognition by radical decomposition,” ACM Transactions on Asian Language Information Processing., vol. 2, no. 1, pp. 27-48, 2003.
[12]L. L. Ma and J. Wu, “On-line handwritten Chinese character recognition based on inter-radical stochastic context-free grammar,” The 2012 International Joint Conference on Neural Networks, Brisbane, QLD, 2012, pp. 1-5.
[13]TensorFlow: Large-scale machine learning on heterogeneous systems, [Online]. Available: https://www.tensorflow.org/. [Accessed: June- 2017].