Abstract:

Most image style transfer tasks are that one model can only correspond to one style, which is inefficient in practical application scenarios. This paper proposes a fast style transfer method of single model and multi-style, which can adapt to any style with only one model. A group of linear changes are used to transform content features and style features respectively, and a combined style loss function is used to reconstruct the image. The Avatar-net method, AdaIN method, Johnson’s fast style transfer method and the linear transformation based style transfer method are analyzed and compared. Using PSNR and SSIM as evaluation indexes, it is concluded that the style transfer method in this paper is better. PSNR reaches 11.591 dB and SSIM reaches 0.499, and the application of this method and video style transfer also have good performance.

Key words: image style transfer, linear transformation, video style transfer

摘要：

多数图像风格迁移任务都是一个模型只能对应一种风格，这在实际应用场景中效率低下，提出一种单模型多风格的快速风格迁移方法，只使用一个模型就可以适应任意风格样式。使用一组线性变化分别对内容特征和风格特征进行转换，使用组合的风格损失函数来重建图像。分析比较了Avatar-net方法、AdaIN方法、Johnson的快速风格迁移方法和基于线性变换的风格迁移方法，并使用PSNR和SSIM作为评价指标，得出提出的风格迁移方法更优，其中PSNR达到了11.591 dB，SSIM达到了0.499，并且将该方法应用于视频风格迁移也有不错的表现。

关键词: 图像风格迁移, 线性变换, 视频风格迁移