| 摘要: |
| :目的 针对大部分风格迁移方法无法在中国水墨画风格领域实现有效迁移的问题,提出一种基于改进扩散
模型采样算法的中国水墨画风格迁移方法,以实现高质量水墨画风格迁移效果。 方法 使用中国水墨画图像数据集
训练去噪 Unet 网络,使得扩散模型获得真实水墨画图像生成能力;设计损失驱动的扩散模型采样算法,先对内容
图像加噪得到包含部分结构信息的含噪数据,然后使用内容损失、风格损失以及语义分离损失,从内容与风格两个
角度引导扩散模型对含噪数据逐步去噪,最终生成风格化结果。 结果 定性与定量对比实验结果证明:所提出的方
法在中国水墨画风格迁移任务上相较于其他任意风格迁移方法有更好的性能,能够生成更高质量的水墨风格化图
像。 消融实验证明:在内容损失中使用两种不同的计算方式以及在风格损失中使用均值方差色彩损失是合理且有
效的。 结论 模型在保持内容结构基本不变的同时生成了合适的真实水墨纹理,为中国水墨画高质量风格迁移提供
了有效方法;生成结果进一步证明:在高质量风格迁移的需求下,使用扩散模型技术有很大的发展空间。 |
| 关键词: 风格迁移 中国水墨画 扩散模型 损失引导 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Chinese Ink Wash Painting Style Transfer Based on Diffusion Model |
|
CHEN Chuan LIU Heng
|
|
School of Computer Science and Technology Anhui University of Technology Maanshan 243000 Anhui China
|
| Abstract: |
| Objective To address the issue that most style transfer methods fail to achieve effective transfer in the domain of
Chinese ink wash painting a Chinese ink wash painting style transfer method based on an improved diffusion model
sampling algorithm is proposed to achieve high-quality ink painting style transfer effects. Methods A denoising UNet
network was trained on a Chinese ink wash painting image dataset to endow the diffusion model with the capability to
generate authentic ink painting images. A loss-driven diffusion model sampling algorithm was designed first noise was
added to the content image to obtain noisy data containing partial structural information then the diffusion model performed
step-by-step denoising of this data guided by content loss style loss and semantic separation loss from both content and
style perspectives ultimately generating the stylized result. Results Qualitative and quantitative comparative experimental
results demonstrated that the proposed method outperforms other arbitrary style transfer methods in Chinese ink wash painting
style transfer tasks producing higher-quality ink wash stylized images. Ablation experiments proved that using two different
computation methods for content loss and employing mean-variance color loss in style loss is reasonable and effective.
Conclusion The model generates appropriate and authentic ink wash textures while preserving the basic content structure
providing an effective method for high-quality Chinese ink wash painting style transfer. The generated results further demonstrate
great potential for the development of diffusion model technology under the demand for high-quality style transfer. |
| Key words: style transfer Chinese ink wash painting diffusion model loss guidance |
