智能交互式复杂纹理特征分析与合成

Texture synthesis is a fundamental problem in image processing. Generally speaking, the goal of texture synthesis is to produce texture images that people required. There are two big families in texture synthesis, the parametric procedural methods and non-parametric exemplar based methods, where recent research mainly focuses on the latter one. A typically exemplar based texture synthesis is given a small texture exemplar, generating a big new texture image that different from the exemplar yet looks like it comes from the same underlying process. During the past two decades, there has been tremendous progress in example-based texture synthesis methods. Researchers have realized that without any control the synthesis cannot guarantee a plausible result for the complex real world textures. Furthermore, texture artists, are rarely interested in simply synthesizing a larger texture from an exemplar; rather, their typical goal is to produce textures for 3D models with a detailed and realistic appearance, which requires precise control over the synthesized result. In this project, we are aiming at controllable texture synthesis by adding guidance channels both for the exemplar and the output texture, where the guidance channel for exemplar is based on texture analysis with the help of users (if needed). Then a set of simple painting tools are provided for users to quickly specify what features or aspects of the texture are important for him/her to control. With these guidance channels, the synthesis can be well controlled. To conclude, we are aiming at creating an interactive system which is simple, intuitive yet powerful for users to produce texture images satisfying the constraints specified by themselves.

纹理合成是图形学、图像处理领域的基础问题之一。广义上讲，纹理合成技术旨在合成符合人们要求的纹理图片。基于样例的纹理合成是其中最重要的方法。典型的基于样例的合成是输入一张小的纹理样图，合成一张大的局部相似但整体不同的新的纹理图像。在过去的20年里，基于样例的纹理合成进展很大，但一方面研究者们逐渐意识到在纹理合成中加入控制是处理真实世界复杂多变的纹理图像的必要手段；另一方面，当前的纹理合成方法对建模师、纹理艺术家们并不实用，艺术家们更希望对合成结果进行精确的控制，而不仅是合成一张大的图像。本项目旨在通过对纹理样图的自动分析，辅以少量人工交互，智能而高效的生成一系列源图像特征引导通道，这些引导通道描述了纹理图像中纹理特征的变化，用户可以参照该引导图，使用一些简单的绘图工具，快速、直观地定义一个新的目标引导通道，并在目标引导图中描述较为重要的控制特征或控制点，从而最终生成符合用户要求的纹理合成图像

围绕智能交互式复杂纹理特征分析与合成，本项目研究了：1）复杂纹理图像交互式分割与控制合成；2）非均匀纹理的特征分析和交互式控制合成；3）基于生成对抗网络的非稳态纹理扩展合成；4）纹理特征分析与纹理迁移在三维模型检索，以及图像任意风格迁移中的应用。通过系列研究，提出多种纹理图像智能分析方法及交互式控制合成方案，使得用户最终能方便合成符和自己要求的纹理结果。项目资助期间，这些研究成果成功发表在国际图形学与视觉顶会，包括Eurographics，SIGGRAPH，CVPR以及NuerIPS，共计4篇（其中2篇为本人一作，2篇为本人通讯作者，两篇一作论文还被SCI源刊收录）；申请专利5项，培养学生5名，圆满完成各项项目预定目标。