143.

[Cover] [Contents] [Index]

Page 226

were not used separately. Instead, for each texture index, the sum of the four directions was used as input to the classifier.

5.6.2.4 Multiplicative autoregressive random field

For each measured block, the corresponding neighbour set N used for model support was defined as (1,0), (0,1), and (−1,−1). Three texture features σ_u², δ_y, and the mean value of entries of vector θ, were selected for the segmentation test.

5.6.3 Segmentation results

The fuzzy c-means clustering algorithm (Chapter 4) was employed to cluster these texture features. The parameters chosen for the fuzzy c-means algorithm were m=2 and number of clusters n=6. The segmentation results are displayed in Figure 5.23.

Results derived from the fractal-based approaches are shown in Figure 5.23a. One can see that Sarker’s algorithm (Equation (5.32)) uses a relatively simple approach, but it covers the image intensity surface very well, and produces higher accuracy than other (multi)fractal estimation methods. Clarke’s method (Equation (5.33)), however, shows a poorer ability to extract texture. The results output by FBM (Equation (5.30)) and Voss’s methods (Equation (5.31)) are similar. The results obtained by using Sarker’s algorithm are next compared with other texture quantisation methods, namely GLCM, MAR, and frequency domain filtering as shown in Figure 5.23b.

As indicated in Figure 5.23b, multifractal methods based on Equation (5.32) do not seem to be very competitive with respect to other texture extraction approaches. GLCM produces the highest accuracy for the noise-free image. However, for noise-fading images, the MAR model seems to show more stable results. Note that texture classifications based on the MAR model use only three features as inputs (i.e. mean value of vector θ, covariance σ^2, and mean δ). These three features appear to characterise texture very well. An added feature of the MAR model is that the computational expense is low. Therefore, the MAR model can be regarded as the most useful tool for texture quantisation.

The GLCM method generally needs much more computational resources than other methods. The computational requirement of the GLCM method is proportional to the number of grey levels used to represent the image. For instance, if the total number of grey levels is 32, the calculation of each index for each pixel and on a specific direction requires a 32×32 matrix. However, if the number of grey levels is 256, the same process will need to operate on a 256×256 matrix, which increases the computational requirements very considerably. Therefore, a direct way of cutting the com

[Cover] [Contents] [Index]