基于非线性滤波和纹理传输的图像风格化

Image Stylization by Non-linear Filtering and Texture Transfer

纹理传输通过将样本图像纹理信息传递给源图像实现风格化.但在处理纹理细节丰富的源图像和复杂样本图像时,现有技术的处理效果差,且不便于用户进行直观的交互操作.为此,提出一种基于非线性滤波的显式纹理传输方法.首先通过相对总变差和L0梯度最小化滤波器对源图像进行去纹理操作,然后通过相对总变差滤波器对样本图像进行纹理提取,最后通过亮度重映射和线性叠加实现纹理传输.实验结果表明,该方法能够在保持源图像结构和色调信息的同时进行纹理替换;与已有方法相比,交互过程更加直观且风格化效果更好.

Texture transfer is a kind of stylization techniques for images, which transfers texture information from an example image to the source image. However, existing methods cannot work well when the source image contains rich textures or the example image is complex, and their interaction is unintuitive. In this paper, we propose a new texture transfer method based on non-linear filtering. Firstly, textures of the source image are removed using relative total variation （RTV） filtering and L0 filtering. Secondly, textures of the example image are extracted using RTV filtering. And finally, the extracted textures are transferred to the source image by illustration remapping and linear superposition. Experimental results show that the pro- posed method can preserve the structure and tone information of the source image while update its textures by those from the example image. Compared with existing methods, our results are more favorable and our interaction is more intuitive.