基于改进CycleGAN的粉笔字书写风格迁移研究
Research on Chalk Calligraphy Style Transfer Based on Improved CycleGAN
陈二开 1李成城 1邬友 2武美玲3
作者信息
- 1. 内蒙古师范大学计算机科学技术学院,呼和浩特 010022
- 2. 内蒙古师范大学科技处,呼和浩特010022
- 3. 吕梁职业技术学院,吕梁 033000
- 折叠
摘要
针对采用循环一致性生成对抗网络(Cycle-consistent generative adversarial network,CycleGAN)进行粉笔字书写风格迁移时,生成的字体存在笔画缺失和模糊等问题,本研究提出改进CycleGAN的粉笔字生成算法.在原始CycleGAN基础上融合自注意力机制来提取粉笔字书法字体的风格特征,并在自注意力机制中使用最大池化和缩放点积来进一步提升模型捕获汉字全局特征的能力.使用相对鉴别生成对抗损失函数改进原网络中的损失函数,以引入先验知识并增强判别器的能力.实验结果表明,使用改进后的模型在学习粉笔字书写风格后生成的字体笔画更加完整,细节更加清晰.
Abstract
While using CycleGAN for chalk calligraphy style transfer,issues like stroke omissions and blurriness occur.To address the issues,an enhanced chalk font generation algorithm based on CycleGAN was proposed.This improvement integrated a self-attention mechanism into CycleGAN,enhancing its ability to capture chalk calligraphy font features.Max-pooling and scaled dot-product attention were used within the self-attention mechanism to capture Chinese character global features better.By employing a relative discriminative generative adversarial loss function,the original network's loss function was enhanced and the discriminator's capability with prior knowledge was bolstered.Experimental results demonstrated that the improved model generates characters with more complete and clearer strokes after learning the chalk calligraphy style.
关键词
粉笔字/风格迁移/CycleGAN/自注意力机制/相对鉴别生成对抗损失函数Key words
Chalk calligraphy/Style transfer/CycleGAN/Self-Attention mechanism/Relative discriminative generative adversarial loss function引用本文复制引用
出版年
2024
NETL
NSTL
万方数据