This preview shows page 1. Sign up to view the full content.
Unformatted text preview: 62 Chapter 11. Solutions: Case Study: Classiﬁed Information Figure 11.3. The images resulting from minimizing D.
is very dependent on the initial guess, but the rather unlikely choice that we made
(all zeros in the green coordinate) gave some of the best results.
Our ﬁrst evaluation criterion should be how the image looks, sometimes called
the “eyeball norm”. In the results for minimizing D, it is harder to diﬀerentiate the
dog from the background. For minimizing R with k = 3, his white fur is rendered as
green and the image quality is much worse than for minimizing D or using k -means.
For k = 4 or k = 5, though, minimizing R yields a good reconstruction, with good
shading in the fur and good rendering of the table legs in the background, and the
results look better than those for minimizing D. (Note that the table legs were not
part of the sample that determined the cluster centers.)
We can measure the quantitative change in the images, too. Each pixel value
xi in the original or the clustered image is a vector with q dimensions, and we can
measure the relative change in the image as
i=1 ||xoriginal − xclustered ||2
i=1 ||xoriginal ||2
i 1/ 2 . This measure is usually smaller when minimizing R rather than D: .363 vs .271 for
k = 3, .190 vs .212 for k = 4, and .161 vs .271 for k = 5. The optimization program
sometimes stops with negative coordinates for a cluster center or no points in the ...
View Full Document
This note was uploaded on 01/21/2012 for the course MAP 3302 taught by Professor Dr.robin during the Fall '11 term at University of Florida.
- Fall '11