Article Preview
Top2. Structural Approach
This approach require that texture have to be composed of basic motifs repeated quasi-regularly (according to some rules) in different orientations on space. They are adapted to images that have a periodic macro-texture.
A texture’s structural description consists firstly of finding the repeated motifs (Texels) then defining their positioning rules.
Figure 1. Example of Texel in texture
Several methods can be regrouped in heuristic methods called classic structural methods that consist to find the primitive and the positioning rules (bottom-up methods).
They rely mainly on signal processing, the topology and geometry. The advantage of these methods is that we can use the classical techniques of segmentation, like the thresholding, edge detection, etc., to find the primitives (motifs) considered as a set of pixels having common homogeneous properties. These properties allow the description of some primitive classes.
There are other methods, called syntactic, use the grammar theory that enables generation of shapes by applying a set of placement rules. However, one texture can be generated by different grammars.
The advantage of the structural approach is that it provides a good symbolic description of the image. Texture primitive can be used to form more complex texture patterns by means of some rules that limit the number of possible arrangements of the primitive.
However, the features generated with these approaches are more useful for synthesis than analysis tasks, although the natural images are more complex than the synthesis images.
Top3. Statistical Appraoch
The texture is seen as a stochastic process, the purpose is to extract the statistical attributes that are non-determinist properties. Thus, texture is described by the statistic of distribution of gray level. This approach is the most adapted to the aperiodic and non-homogeneous texture.
There are many measure orders depending on the number of pixels used in the statistic analyze, when one pixel is used, we call this first order measure. Thus, in the second order (two pixels are used) we analyze the distributions and relationships between a pixel and one of its neighbors (pixel pair) (Haralick, 1979).This order has proved its efficiency relatively to the human perception (Julesz, 1975). The most popular statistical features in this order can be extracted from the method called the co-occurrence matrix (Haralick, 1979); this last was used in the biomedical-images (Lerski, 1993) (Strzelecki, 1995) and has proved better results than the wavelet packets (a transform-based technique) in the domain of texture classification (Valkealathi, 1998).
There are many methods, based in this approach, that have been invented. We will discuss about the most common ones.