Comparison with Related Synthesis Methods

Today's mainstream texture synthesis methods [29,104,65,59,28,78] are based on non-parametric techniques, which are reviewed in detail in Section 5.3. Similar to bunch sampling, these methods also use image signals retrieved from the training image to grow synthetic textures. But these methods adopt a very different approach to avoid the time-consuming identification of a probability model. These methods simply skip the entire stage of building texture models from global image statistics, but instead use non-parametric techniques, e.g., neighbourhood matching, to replicate texture features. Without a texture model, non-parametric sampling methods are unaware of the global structure of a texture, so they have difficulty in determining the size or shape of local neighbourhood that suits the current texture. In most of these methods, user intervention is required to provide a proper local neighbourhood for the algorithms. Another issue is that the synthesis tends to accumulate error or grow garbage, because a texture is synthesised either pixel by pixel or block by block, in a sequential manner. Because they generate textures based primarily on local constraints, non-parametric sampling techniques might outperform the bunch sampling in reproducing coherent local texture features. But their weakness is especially in replicating the global characteristics of a regular texture.

Figure 7.7: A synthesis result of bunch sampling (BS) compared with others (part of the images are taken from [66]).

Input	Lattice	BS	Efros&Leung [29]
Quilting [28]	Wei&Levoy [104]	Liu [66]	Graph Cut [59]

In its basic idea, the method for synthesising near-regular textures proposed in [67] is similar to bunch sampling. But in the former approach, periodicity structures in regular textures are recovered from the translation symmetries of autocorrelation functions. Since they are based on statistics of pairwise signal products over the clique families, autocorrelation functions describe interaction structure in a less definite way than the general statistics of pairwise signal co-occurrences in a generic MGRF model. Another difference from bunch sampling is in that the method uses large image tiles as construction units for texture synthesis, i.e. each tile is cut from a training image and then placed into a synthetic texture in line with the translation lattice. But this approach requires to blend the overlapping regions at seams to avoid visual disruption [67].

Figure 7.7 shows the comparison of synthesis results for a `mesh' texture among the bunch sampling, non-parametric sampling [29,104,59,28], and Liu's algorithms [67]. Although all the methods produce similar synthetic textures, philosophies are different between bunch sampling and others. An interesting fact is that only bunch sampling discovers the periodicity structure hidden in the texture. The translation lattice derived by the bunch sampling is shown by the `lattice' image in Fig 7.7. Therefore, the bunch sampling sees this texture as a perfect regular texture, while other methods, focusing on local neighbourhood, treat the texture as a weakly-homogeneous one. Since a texture is a mixture of global and local features, neither model-based bunch sampling nor non-parametric sampling is adequate for complete texture modelling. The future development of texture analysis and synthesis would focus on incorporating both the global and the local characteristics of a texture into a better image model.