Image Processing Reference
In-Depth Information
Fig. 16.9. ( Reading direction ) In pseudocolor, the images represent the four most significant
texture features of P3 after dimension reduction
were three frequencies in the decomposition, at each of which I 11 (real), I 20 (com-
plex), I 40 (complex) were computed. Taking the real and the imaginary parts of the
complex features, one thus obtains 15 real features. Furthermore, the dimension of
this space was reduced to five by projecting these features to the five most significant
eigenvectors from a KL transform, see Chap. 15. Consequently, the clustering and
the boundary refinement procedure here were fed with 5 dimensional features, of
which are 4 are shown in Fig. 16.9. We required N c =7
9 classes from the clus-
tering with L =3, i.e., the clustering had to be performed on a 16 × 16 image, and
the correct patch identities (up to a texture name permutation) as well as reasonably
accurate boundaries could be obtained in an unsupervised manner, i.e., there was
no training. The boundaries are nondeterministic i.e., they show small variations be-
tween the runs because the prototypes display small variations. This depends on the
−−
Search WWH ::




Custom Search