Acta Scientiarum Naturalium Universitatis Pekinensis ›› 2024, Vol. 60 ›› Issue (6): 989-1000.DOI: 10.13209/j.0479-8023.2024.088

Previous Articles     Next Articles

An Image Segmentation Based Technology for Intelligent Character Creation

JIANG Jianbin1,2, HUANG Song1, WU Jiangguo1,†   

  1. 1. College of Engineering, Peking University, Beijing 100871 2. Beijing Founder Electronics Co., Ltd, Beijing 100086
  • Received:2024-01-08 Revised:2024-06-28 Online:2024-11-20 Published:2024-11-20
  • Contact: WU Jiangguo, E-mail: j.wu(at)pku.edu.cn

基于图像分割技术的智能造字研究

蒋建斌1,2, 黄松1, 吴建国1,†   

  1. 1. 北京大学工学院, 北京 100871 2. 北京北大方正电子有限公司, 北京 100086
  • 通讯作者: 吴建国, E-mail: j.wu(at)pku.edu.cn
  • 基金资助:
    国家自然科学基金(72171003, 71932006)资助

PDF

Abstract:

Currently, the intelligent character creation technology based on style learning generates fonts with low similarity to the user’s handwritten style, and the method based on Graphics Processing Unit (GPU) style transfer is expensive. To solve the problems above, a new intelligent character creation method is proposed using deep learning and image segmentation technology, which can maintain a highly similar style while meeting the personalized needs and reducing the generation cost. Using DeepLab v3+ technology, 775 font images input by users are filtered by a data quality evaluation model and are split to components by image segmentation models. Then, components are finely adjusted and noise is removed, and finally TrueType fonts are generated after vectorization. Compared with existing technologies, this method significantly improves the similarity and reduces the cost, and can effectively meet the personalized customization needs of users.

Key words: intelligent character creation, DeepLab v3+, data quality evaluation model, image segmentation model; Truetype

摘要：

目前, 基于风格学习的智能造字技术生成的字体与用户手写风格相似度低, 基于GPU (graphics processing unit)风格迁移的方法成本高昂。为解决上述问题, 利用深度学习和图像分割技术, 提出一种新型智能造字方法, 在保持高度相似风格的同时, 满足用户个性化需求, 并降低成本。采用DeepLab v3+技术, 用户输入的775个字体图像经过数据质量评估模型筛选后, 通过图像分割模型进行部件拆分, 然后精细地调整部件并去除噪点, 最终矢量化后生成TrueType字体。与现有技术相比, 该方法能够显著地提升相似度并降低成本, 可以有效地满足用户个性化定制需求。

关键词: 智能造字, DeepLab v3+, 数据质量评估模型, 图像分割模型, TrueType

Copyright © Acta Scientiarum Naturalium Universitatis Pekinensis, All Rights Reserved.
Powered by Beijing Magtech Co., Ltd.