18.11.2014 Views

Pit Pattern Classification in Colonoscopy using Wavelets - WaveLab

Pit Pattern Classification in Colonoscopy using Wavelets - WaveLab

Pit Pattern Classification in Colonoscopy using Wavelets - WaveLab

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

4.3 Structure-based classification<br />

s<strong>in</strong>ce d Ti ,T i<br />

is just the distance of a quadtree of an image to itself, which must by def<strong>in</strong>ition<br />

of a metric always be 0. Therefore D is<br />

⎛<br />

⎞<br />

0 d T1 ,T 2<br />

. . . d T1 ,T n<br />

d T2 ,T 1<br />

0 . . . d T2 ,T n<br />

D = ⎜<br />

⎝<br />

.<br />

. . ..<br />

⎟ . ⎠<br />

d Tn,T1 d Tn,T2 . . . 0<br />

This matrix then conta<strong>in</strong>s the quadtree distances (similarities) between all possible pairs<br />

of images out of the tra<strong>in</strong><strong>in</strong>g image set.<br />

For each class c the matrix conta<strong>in</strong>s a submatrix D c which conta<strong>in</strong>s the distances between<br />

all images belong<strong>in</strong>g to class c. Therefore D can also be written as<br />

⎛<br />

⎞<br />

0 d T11 ,T 12<br />

. . . d T11 ,T 1n<br />

. . . d T11 ,T C1<br />

. . . d T11 ,T Cn<br />

d T12 ,T 11<br />

0 . . . d T12 ,T 1n<br />

. . . d T12 ,T C1<br />

. . . d T12 ,T Cn<br />

. . . .. . . . . .<br />

d<br />

D =<br />

T1n ,T 11<br />

d T1n ,T 12<br />

. . . 0 . . . d T1n ,T C1<br />

. . . d T1n ,T Cn<br />

. . . . . .. . . .<br />

d TC1 ,T<br />

⎜<br />

11<br />

d TC1 ,T 12<br />

. . . d TC1 ,T 1n<br />

. . . 0 . . . d TC1 ,T Cn<br />

⎝<br />

. ⎟<br />

. . . . . . .. . ⎠<br />

d TCn ,T 11<br />

d TCn ,T 12<br />

. . . d TCn ,T 1n<br />

. . . d TCn ,T C1<br />

. . . 0<br />

where d Tij ,T kl<br />

is the distance between the j-th image (quadtree) of class i and the l-th image<br />

(quadtree) of class k.<br />

Figure 4.5(a) shows an example distance matrix for the two-class case. This example<br />

shows a perfect distance matrix with very small <strong>in</strong>tra-class distances (black boxes) and very<br />

high <strong>in</strong>ter-class distances (white boxes). In reality however, a distance matrix is more likely<br />

to look like the one illustrated <strong>in</strong> figure 4.5(b), s<strong>in</strong>ce the trees for different images from a<br />

specific class hardly will be totally identical <strong>in</strong> real-world applications.<br />

As can been seen <strong>in</strong> figure 4.5 and follow<strong>in</strong>g from equation (4.40) the distance matrices are<br />

all symmetric.<br />

Hav<strong>in</strong>g calculated the matrix D the next task is to f<strong>in</strong>d the centroid of each class. A<br />

centroid of a class c <strong>in</strong> this context is the tra<strong>in</strong><strong>in</strong>g image out of I Tc which has the smallest<br />

average distance to all other images of class c. In other words, it is now necessary to f<strong>in</strong>d the<br />

image of class c to which all other images of class c have the smallest distance <strong>in</strong> respect to<br />

the chosen quadtree distance metric. After the centroids of all classes have been found the<br />

classification process can now take place.<br />

To classify an arbitrary image first of all the image is decomposed us<strong>in</strong>g the wavelet<br />

packets transform to get the respective decomposition tree. It is important to note that for<br />

classification the same cost function has to be used for prun<strong>in</strong>g the decomposition tree as<br />

was used dur<strong>in</strong>g the tra<strong>in</strong><strong>in</strong>g phase for all tra<strong>in</strong><strong>in</strong>g images. Then the distances of the result<strong>in</strong>g<br />

decomposition tree to the decomposition trees of the centroids of all possible classes are<br />

55

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!