wuyongf / font2font

An enhanced zi2zi project with word-oriented data augmentation, feature combination, and transfer learning.

Font2Font

Yun-Chen Lo, Yi-Ren Chen, Jerry Zhang Jian (NTHU EE)

An enhanced zi2zi project [1] with word-oriented data augmentation, feature combination, and transfer learning.

This repo is based on zi2zi project with incremental boost inspired on recent research and our observation on Chinese Characters.

Folder Description

Font2Font
|	README.md
└──	src/
|	└── zi2zi/					# original zi2zi with data aug options
|	└── zi2zi_hir/					# + combine levels of features
|	└── zi2zi_hir_dis/				# + increase discriminator complexity
|	└── zi2zi_hir_morefilter/			# + alternative ways of complexity boost
|	PastDiscussion/
|	PriorWork/
|	Report/
|	└── Interim/
|	└── Final/

Code Running Example

Train zi2zi 
change directory to src/zi2zi;
$ make all // prepare data, train, and evaluate on demo word

Dependency

• tensorflow 0.8
• python 3.x

Major contributions

Word-oriented data augmentation

Combine different level feature

Transfer Learning

Difficulty-aware adjustment

Result

Future Work

Add code or tools to generate compatible Chinese font (.ttf)

Add handwriting dataset and tools for further application exploration