i
SI
€
A APUD
Fig. 9 PV1 image, size 128 z 256 consisting of a num-
ber of signals in different sizes and orientations. The image
is heavily corrupted by a computer generated o,, = 50 noise
field, yielding a SNR = 4. The lower image shows the
outlines of the detected regions after passing the entire pro-
cedure.
The parameters of the predictor are: w = 1.5, trunca-
tion factor — 0.01, resulting in a half plane filter size of 6
x 11. The confidence levels are, Stage I: 95%, Stage Ila:
99.9% and Stage IIc: 99.996. In Stage IIb regions smaller
than 3 pixels are removed; c, — 1.
Stage I yields 239 regions, represented in fig. 8d, where
each region is shown with different grey value. Stage Ila
merges 230 of the 239 regions. In Stage IIb two regions are
smaller than 3 pixels. They are merged with their most
similar neighbours. In Stage IIc no insignificant regions
are traced. The result of the entire procedure, where each
region has received a grey value, is shown in fig. 8e. Fig. 8f
displays the region outlines.
Smoothing of the SB images corrupted by o, = 10 or
20 doesn't improve the final result, since the predictor has
already a smoothing effect. Heavy noise o, — 50 needs
smoothing in Stage I and Stage II. Best performance in
Stage I shows the extended Kuwahara filter and next best
the Gaussian filter. The conditional filter shows worst be-
700
havior; the median and box filter lie in between. Our ex-
periments confirmed out earlier considerations that the me-
dian filter in Stage II gives best results, since it removes the
tail parts of the noise enabling a more reliable decision on
merging.
The PV1 Images (fig. 9a) shows signals in a number
of sizes and orientations, contaminated by a o, — 50 noise
field. Predictor parameters: see SB image. The results
of c, — 10 and o, — 20 images are good, even without
previous smoothing. Smoothing of the 0, = 50 image by
extended Kuwahara or Gaussian, showed, according our
previous observation, the best results in Stage I. In Stage
II the median filter shows best performance.
Experiments on setting the confidence levels yielded:
Stage I: 9596, Stage IIa: 99.996 and Stage IIc: 9596, as best
options. Fig.9b shows the outlines of the regions obtained,
where the extended Kuwahara in Stage I and the median
filter in Stage II is applied, using the above best options.