Scholarlink

<p>The field of artificial intelligence-generated content (AIGC) has made significant progress in handwriting generation, including complex scripts such as Chinese. Previous offline Chinese handwriting generation methods focus on isolated character image generation, without addressing the task of offline Chinese handwritten text line generation. This paper proposed a novel multi-level style control method for generating Chinese handwritten text lines, which consists of two key steps: style-controlled character image generation and text-line layout generation. For style-controlled character image generation, we implement a pre-training method utilizing multi-modal radical-level contrastive loss to align image features with radical embeddings in encoders. Additionally, a multi-level style representation control is achieved through multi-modal feature aggregation. We also propose a style-consistent text-line layout generation scheme by using prompt engineering with a large language model. Experimental results demonstrate that our method achieves comparable or even better performance in character image generation compared to diffusion model-based methods, while also delivering faster generation speeds. By incorporating text-line layout generation, the generated text-line samples can be effectively used for training handwriting recognition models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-level Style Control for Chinese Handwriting Generation

  • 期刊论文

  • 发表日期：

    2025年6月7日

查看全文
  • Gang Yao,
  • Liangrui Peng,
  • Zhiyu Li,
  • Tianqi zhao,
  • Kemeng Zhao,
  • Ning Ding,
  • Yao Tao

摘要

The field of artificial intelligence-generated content (AIGC) has made significant progress in handwriting generation, including complex scripts such as Chinese. Previous offline Chinese handwriting generation methods focus on isolated character image generation, without addressing the task of offline Chinese handwritten text line generation. This paper proposed a novel multi-level style control method for generating Chinese handwritten text lines, which consists of two key steps: style-controlled character image generation and text-line layout generation. For style-controlled character image generation, we implement a pre-training method utilizing multi-modal radical-level contrastive loss to align image features with radical embeddings in encoders. Additionally, a multi-level style representation control is achieved through multi-modal feature aggregation. We also propose a style-consistent text-line layout generation scheme by using prompt engineering with a large language model. Experimental results demonstrate that our method achieves comparable or even better performance in character image generation compared to diffusion model-based methods, while also delivering faster generation speeds. By incorporating text-line layout generation, the generated text-line samples can be effectively used for training handwriting recognition models.