欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 图像与视频处理 • 上一篇    下一篇

基于生成式对抗神经网络的手写文字图像补全

  

  1. (1. 东华理工大学经济与管理学院，江西 南昌 330013； 2. 华南理工大学电子与信息学院，广东 广州 510641； 3. 牛津大学数学研究所，牛津 OX26GG)
  • 出版日期:2019-10-31 发布日期:2019-11-06

Handwritten Character Completion Based on  Generative Adversarial Networks

  1. (1. School of Economics and Management, East China University of Technology, Nanchang Jiangxi 330013, China; 
    2. School of Electronic and Information Engineering, South China University of Technology, Guangzhou Guangdong 510641, China; 
    3. Mathematical Institute, University of Oxford, Oxford OX26GG)
  • Online:2019-10-31 Published:2019-11-06

PDF (PC)

207

摘要： 手写文字图像补全是图像补全问题中一个重要研究分支，其难点在于图片中具有 无约束书写风格的文字的结构关系补全。为了模拟实际中复杂和困难的应用情景，在图像补全 研究工作的启发下，针对大类别、小样本、多风格、未知语种等复杂情况下进行手写象形文字 图像补全。采用全局和局部一致性保持的生成式对抗神经网络(GLC-GAN)。在大类别多风格的 手写文字图像补全中，补全图片往往因可能的补全候选很丰富而导致补全区域模糊不清。为此， 提出两级补全系统：第一级粗补全模块考虑文字结构的完整性，第二级细补全模块实现文字的 清晰化、细致化。通过在大类别手写汉字数据库 CASIA-HWDB1.1 上的实验，验证了该两级系 统的有效性，同时分析系统在不同书写风格和不同缺失区域情况下的补全效果。

关键词: 生成式对抗网络, 手写文字, 图像补全, 结构补全, 自监督学习

Abstract: Handwritten character completion is an important research topic in image completion. Its challenge comes from the completion of the structural relationships in handwritten characters with unconstrained handwritten styles. To simulate the complicated and difficult situations in the real-world applications, the paper focuses on handwritten pictographic characters with large category, small sample size, multiple unconstrained handwritten styles, and unknown language (i.e., with no access to the class label of each character). Inspired by the progress in natural image completion, the generative adversarial network with global and local consistency was leveraged to achieve handwritten character completion. Under the circumstances of large category and various writing styles, the completion areas of character completion suffer from low-fidelity because of the large number of potential completion candidates. To solve this problem, a two-stage character completion system was proposed: the first stage is coarse-grained completion module ensuring the completeness of the character; the second stage is fine-grained completion module improving the sharpness and details of characters. Extensive experiments were conducted on CASIA-HWDB1.1 to validate the effectiveness of the two-stage system and analyze the completion performance under different writing styles and different conditions of missing area.

Key words: generative adversarial network, handwritten character, image completion, structure completion, unsupervised learning