159.

[Cover] [Contents] [Index]

Page 240

Figure 6.4 Metropolis et al. (1953) algorithm for creating a texture image.

The procedure illustrated in Figure 6.4 allows the generation of a unique texture pattern for an image. If an image with multitexture patterns is required, then a hierarchical model can be applied. A hierarchical model generally involves two levels. The higher level uses a MLL model with isotropic potential parameter (i.e. unique β) to generate a blob-like image, which is sometimes called a region process. At the lower level, each blob region is filled with different texture patterns generated by the MLL model using anisotropic potential parameters (i.e. different β for different orientations). The resulting image will therefore contain multitexture patterns. At the higher level, one can sometimes use a predefined (or digitised) image to replace the region process. At the lower level of the hierarchical model, one may also use independent noise instead of texture patterns to fill each blob region. Figure 6.6 illustrates the construction of multiple texture images using a hierarchical model.

Several studies (Derin and Cole, 1986; Derin and Elliott, 1987; Won and Derin, 1992) use images generated by the hierarchical model to test the robustness of image segmentation algorithms. We can regard the higher level image of a hierarchical model as representing the desired segmentation, which is contaminated by texture or noise generated at the lower level of the hierarchical model. If the classification is based on Bayesian formula, one can adopt a smooth assumption to model prior p.d.f. (Section 6.1.4), while the conditional p.d.f. of the data given class label w can be

[Cover] [Contents] [Index]