基於深度學習的中文字型生成之研究與實作

A Study and Implementation of Chinese Font Generation Based on Deep Learning

稅顯堯(Hsien-Yao Shui)

指導教授 ： 張智星

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202103746

摘要

由於存在大量字符，因此生成中文字體庫是一項耗時的工作。以繁體中文設計為主需要開發超過 50,000 個字符，即使是比較小的 BIG5 集也將由 13,053 個中文字組成。當手寫中文字體具有各種不同的筆劃和部首結構樣式時，開發個性化的中文字體庫會更加困難。大多數個性化的中文字體產品都要求使用者先提供超過 1,000 個以上的手寫字，或者只能選擇有限風格的字體，無法滿足使用者多樣化的需求。近年來，許多研究提出了自動生成中文字體的AI 和計算機圖形系統，但是在輸出質量上仍然存在缺陷。在一方面基於筆劃的字體合成方法在構建筆劃提取資料庫方面遇到挑戰，繁瑣的人工的微調仍然不可避免。在另外一方面是基於學習的方法出現筆劃模糊，不完整或不正確的筆劃等常見問題。大部份的學習方法都要求用戶輸入以上 500 個字符，仍然缺乏實用性。在這項工作中，我們提出了一個兩階段的中文字體生成系統，該系統僅需要用戶編寫 25 個手寫字符。我們的系統在第一階段結合了字體樣式分類網絡，並在第二階段結合 CycleGAN 和骨架轉換網絡。實驗結果證明，與其他基於深度學習的方法相比，我們的方法具有更好的效能，除了產生風格更相近的個性化中文字體，也完美地解決了常見的中文字形生成的品質問題。

關鍵字

中文字形生成 ； 風格轉換 ； 字體合成 ； CycleGAN

並列摘要

Generating a Chinese font library is a time-consuming task due to the large number of characters. There are over 50,000 characters to be developed in Traditional Chinese and even the smaller BIG5 set is composed of 13,053 Chinese characters. It is more difficult to develop a personalized Chinese font library while handwriting of Chinese has various different styles of stroke and radical structures. Most personalized Chinese font creator products are human creativity required or base on the commercial design templates. In recent years many AI and Computer Graphics system are proposed for automatic generation of Chinese fonts, however, there are still drawbacks in the output quality. Existing stroke-based font synthesis methods have challenges to build the stroke extraction database and human finetuning are still unavoidable. All learning-based methods have common problems of blurred, incomplete or unreasonable strokes and they require 500 characters above from user input. In this work we propose a two-stage Chinese font generation system, which requires only 25 characters written by the user. Our system combines the font style classification network in the first stage, and the CycleGAN with joint skeleton transfer training in the second stage. Experimental results verify the better performance of our method compared to other deep learning based approaches and resolve the common problems.

並列關鍵字

Chinese font generation ； style transfer ； font synthesis ； CycleGAN

參考文獻

[1] ShowYourWords, WRITES Co. https://www.writes.com.tw/.

[2] Yuchen Tian, “Rewrite: Neural Style Transfer For Chinese Fonts.” Retrieved Nov 23, 2016 from https://github.com/kaonashi-tyc/Rewrite, 2016.

[3] Yuchen Tian, “zi2zi: Master Chinese Calligraphy with Conditional Adversarial Networks.” Retrieved Jun 3, 2017 from https://github.com/kaonashi-tyc/zi2zi, 2017.

[4] ZC119, “Generating handwritten Chinese characters using CycleGAN.” https://github.com/ZC119/Handwritten-CycleGAN.

[5] Alex Zhong, “High Performance Offline Handwritten Chinese Character Recognition Using GoogLeNet and Feature Maps.” https:// github.com/ zhongzhuoyao/HCCR-GoogLeNet.