Computational Vision and Medical Image Processing IV

86

6. Check that there are still points in the cloud. If

so, go back to the pt.2

7. Select the first k clusters determined by this

algorithm. Their centroids are final starting cen-

tres for K-means clustering.

The MM initialization permits to get the red

letter in the image during the quantisation onto

4 colours (see Figure 2e). For all the considered

initializations we calculated a colour error for the

red letter: ΔE (DE) = 77, ΔE (SD) = 60 and ΔE

(MM) = 11. These results demonstrate the superi-

ority of MM initialization.

Figures 3–6 illustrates subsequent eliminations

of outlier clusters from the cloud of points and

helps to understand the algorithm. Here, the third

step of MM (see Fig.3) has particular importance,

because a centre of red ...