142.

[Cover] [Contents] [Index]

Page 225

Figure 5.22 The behaviour of D(q) with binomial multiplicative process.

approximately the same. Therefore, the values q=0, q=3, q=−4 were used in our experiments.

5.6.2.2 Fourier power spectrum

Two kinds of features extracted by applying W and R filters (Equations (5.38) and (5.39)) are used for the purposes of this example. The first is the proportion of the total power spectrum contained in the W filter, i.e. [W_θ1,_θ2/ P(u, v)]. Only horizontal and vertical directions within a predefined angular range of 45° (i.e. ±22.5°) were used. The second feature is the proportion of the power spectrum contained in the R filter, i.e. [R_r1,_r2/ P(u, v)]. Since both features are sensitive to orientation (i.e. [W_θ1,_θ2/P(u,v)]) only, or to frequency (i.e. [R_r1,_r2/P(u, v)]) only, it is better to use both filters together to perform texture segmentation. Note that there is no clear rule concerning the choice of a suitable range interval (r₁, r₂) for the R filter. The range was therefore decided by trial and error methods.

5.6.2.3 Grey level co-occurrence matrix

Four texture features, ASM, Con, Ent, and IDM were derived from each subimage for the H, V, LD, RD directions with distance d=1. In order to avoid the angular effect, in which the same texture pattern may spread out in different directions, the feature indices generated for each direction

[Cover] [Contents] [Index]