Hybrid texture synthesis

Patch-based texture synthesis algorithms produce reasonable results for a wide variety of texture classes. They preserve global structure, but often introduce unwanted visual artifacts along patch boundaries. Pixel-based synthesis algorithms, on the other hand, tend to blur out small objects while maintaining a consistent texture impression, which in return doesn't necessarily resemble the input texture. In this paper, we propose an adaptive and hybrid algorithm. Our algorithm adaptively splits patches so as to use as large as possible patches while staying within a user-defined error tolerance for the mismatch in the overlap region. Using large patches improves the reproduction of global structure. The remaining errors in the overlap regions are eliminated using pixel-based re-synthesis. We introduce an optimized ordering for the re-synthesis of these erroneous pixels using morphological operators, which ensures that every pixel has enough valid (i.e., error-free) neighboring pixels. Examples and comparisons with existing techniques demonstrate that our approach improves over previous texture synthesis algorithms, especially for textures with well-visible, possibly anisotropic structure, such as natural stone wall or scales.