構建一個完整的個性化中文字體庫，相較於其他的語言的字體設計上，是相
對困難的任務。在中文字裡，相較於其他語言，如英文由 26 個字母組成、西班牙
文由 27 個字母，在繁體中文字中常用字數就超過了一萬三千個字，而且每個字
都會是以多個部件進行組合、排列而成。因此如何克服大量的字數及更有效率的
設計字型，解決需要耗費相當多的時間及專業人力的問題，是本論文所討論的重
點。

隨著深度學習興起，自動生成由大量結構複雜的字形組成的中文字體仍然是
一個具有挑戰性的問題。所以過往字體生成的方法，目標都在於如何擴增與強化
特徵來達到更好的效果，因此會使用大量字集、前處理、多模組的訓練的方式來
達成。在使用大量字集上，需要透過大量的中文字樣本集作為訓練，提升模型的
泛化能力，藉此提高風格轉換結果。在前處理的方式上，會先讓模型先學習其他
字體的特徵分布，使得學習新字體能夠在前處理過的模型上擁有更好的表現。而
在使用多模組的方式上，需透過多個模組進行特徵抽取，最後進行合併達成特徵
擴增，進而獲得更佳的效果。因此過往字體生成的方法需要蒐集更多的樣本集與
更複雜的模型架構方式下來達成。

我們為了解決上述等問題，我們提出一個在特徵抽取與強化上更有效的模型架構
方法來達成。該架構將 Unet 上的 skip-connection 特點與 DenseNet 的
DenseBlock 概念結合成一個端到端的單一模型，強化模型的特徵抽取能力，並利
用三個面向的損失函數來進行模型修正。實驗評估結果顯示，基於 FID 的評分標
準下比較當前的方法，我們提出的 Dense-UFont 在與 zi2zi 相比下，能夠將字
數降低 40%，與 pix2pix 相比，能夠將字數降低 60%。我們的方法能夠更簡單且
更有效的轉換目標風格中文字體，並且在少量的樣本集中上，能獲得較佳的結果。

Building a complete personalized Chinese font library is a relatively difficult task compared to font design for other languages. In Chinese characters, compared with other languages, such as English consists of 26 letters and Spanish consists of 27 letters, there are more than 13,000 characters commonly used in traditional Chinese characters, and each character will be Combining and arranging with multiple parts. Therefore, how to overcome a large number of words and design fonts in more efficiently way also solve the problem that requires a lot of time and professional manpower, which is the focus of this paper.

With the rise of deep learning, automatic generation of Chinese fonts consisting of a large number of complex glyphs remains a challenging problem. In the past, the goal of font generation methods was how to expand and strengthen features to achieve better results. Therefore, a large number of character sets, preprocessing and multi-module training are used to achieve this. When using a large number of character sets, it is necessary to train through a large number of Chinese character sample sets to improve the generalization ability of the model, thereby improving the results of style transfer. In the preprocessing method, the model will first learn the feature distribution of other fonts, so that learning new fonts can have better performance on the preprocessing model. In the method of using multiple modules, it is necessary to perform feature extraction through multiple modules, and finally merge them to achieve feature amplification, so as to obtain better results. Therefore, the previous methods of font generation need to collect more sample sets and more complex model architectures to achieve.

In order to solve the above problems, we propose a more effective model architecture method in feature extraction and reinforcement. The architecture combines the skip-connection feature on Unet and DenseNet's DenseBlock concept into an end-to-end single model, enhancing the model's feature extraction capabilities, and using three oriented loss functions for model correction. The experimental evaluation results show that compared with the current methods under the FID-based scoring standard, our proposed Dense-UFont can reduce the word count by 40\% compared with zi2zi and 60\% compared with pix2pix. Our method can convert target-style Chinese fonts more simply and efficiently, and achieve better results on a small sample set.

[1] “justwrite,” 2014. http://www.justwrite.tw/.
[2] L. A. Gatys, A. S. Ecker, and M. Bethge, “Image style transfer using convolutional
neural networks,” in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 2414–2423, 2016.
[3] Y. Taigman, A. Polyak, and L. Wolf, “Unsupervised cross-domain image genera-
tion,” arXiv preprint arXiv:1611.02200, 2016.
[4] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation
using cycle-consistent adversarial networks,” arXiv preprint, 2017.
[5] S. Baluja, “Learning typographic style,” arXiv preprint arXiv:1603.04000, 2016.
[6] J.-c. Feng, C. K. Tse, and Y. Qiu, “Wavelet-transform-based strategy for generating
new chinese fonts,” in Proceedings of the 2003 International Symposium on Circuits
and Systems, 2003. ISCAS’03., vol. 4, pp. IV–IV, IEEE, 2003.
[7] S. Xu, H. Jiang, T. Jin, F. C. Lau, and Y. Pan, “Automatic generation of chinese
calligraphic writings with style imitation,” IEEE Intelligent Systems, vol. 24, no. 02,
pp. 44–53, 2009.
[8] Q. Li, J.-P. Li, and L. Chen, “A bezier curve-based font generation algorithm for
character fonts,” in 2018 IEEE 20th International Conference on High Performance
Computing and Communications; IEEE 16th International Conference on Smart
City; IEEE 4th International Conference on Data Science and Systems (HPCC/
SmartCity/DSS), pp. 1156–1159, IEEE, 2018.
[9] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected
convolutional networks,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 4700–4708, 2017.
[10] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for
biomedical image segmentation,” in International Conference on Medical image
computing and computer-assisted intervention, pp. 234–241, Springer, 2015.
28
[11] B. Zhou, W. Wang, and Z. Chen, “Easy generation of personal chinese handwritten
fonts,” in 2011 IEEE International Conference on Multimedia and Expo, pp. 1–6,
July 2011.
[12] Z. Lian, B. Zhao, and J. Xiao, “Automatic generation of large-scale handwriting
fonts via style learning,” in SIGGRAPH ASIA 2016 Technical Briefs, SA ’16, (New
York, NY, USA), pp. 12:1–12:4, ACM, 2016.
[13] A. Myronenko and X. Song, “Point set registration: Coherent point drift,” IEEE
transactions on pattern analysis and machine intelligence, vol. 32, no. 12, pp. 2262–
2275, 2010.
[14] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with con-
ditional adversarial networks,” in Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 1125–1134, 2017.
[15] S. Azadi, M. Fisher, V. Kim, Z. Wang, E. Shechtman, and T. Darrell, “Multi-content
gan for few-shot font style transfer,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, vol. 11, p. 13, 2018.
[16] Y. Jiang, Z. Lian, Y. Tang, and J. Xiao, “Dcfont: an end-to-end deep chinese font
generation system,” in SIGGRAPH Asia 2017 Technical Briefs, pp. 1–4, 2017.
[17] C. Wen, Y. Pan, J. Chang, Y. Zhang, S. Chen, Y. Wang, M. Han, and Q. Tian,
“Handwritten chinese font generation with collaborative stroke refinement,” in Pro-
ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision,
pp. 3882–3891, 2021.
[18] Y. Tian, “zi2zi: Master chinese calligraphy with conditional adversarial networks,”
2017.
[19] B. Chang, Q. Zhang, S. Pan, and L. Meng, “Generating handwritten chinese charac-
ters using cyclegan,” in 2018 IEEE Winter Conference on Applications of Computer
Vision (WACV), pp. 199–207, IEEE, 2018.
29
[20] Y. Jiang, Z. Lian, Y. Tang, and J. Xiao, “Scfont: Structure-guided chinese font gener-
ation via deep stacked networks,” in Proceedings of the AAAI conference on artificial
intelligence, vol. 33, pp. 4015–4022, 2019.
[21] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained
by a two time-scale update rule converge to a local nash equilibrium,” in Advances
in neural information processing systems, pp. 6626–6637, 2017.