PE&RS June 2015 - page 466

Figure 5a and 5b show that when NF0 is set small, i.e., 1
and 3, the changes of scale threshold (Sk) are significant, mak-
ing NR change significantly accordingly. However, when NF0
is set as 6, 10, and 20, the changes of Sk and NR are smoother
and slower than those with small NF0. Moreover, the changing
rate for the three large NF0 is similar with each other. The dif-
ferent changing rates of Sk and NR result in different numbers
of segmentation scales. Consequently, the coverage of mean-
ingful segmentation scales would be different.
For T1, the segmentations scales with approximately 150
to 1,100 regions are viewed as meaningful candidates because
various objects can be represented well within this range
through visual analysis. Given a decreasing rate β , the number
of scales within the meaningful range produced by setting dif-
ferent NF 0 is shown in Table 2. When NF 0 is equal to or larger
than 6, the number of meaningful scales is similar for both the
LMM and LBM methods, but it is decreased significantly when
NF0 is set as 3 and 1, leading to the limited candidate scales.
Combining Figure 5b and 5c, we can see that the meaning-
ful scales are controlled by the normalized factors approxi-
mately ranging from 4 to 5.5 when is NF0 20, 10, and 6. Hence,
the large NF0 can capture the sensitive normalized factors and
make the scale parameter change smoothly to cover meaning-
ful segmentation scales. On the other hand, when NF0 is 3 or
1, since it is smaller than the sensitive normalized factors,
the scale parameter is increased too fast, leading to limited
numbers of candidates to cover meaningful scales.
According to the above analysis, we can know that if NF0
is set smaller than the sensitive normalized factors, it cannot
produce enough candidates to cover meaningful segmentation
scales. The number of meaningful scales is similar when setting
different NF0 if they are larger than the sensitive normalized fac-
tors. However, the sensitive normalized factors for different im-
ages would be various. We set NF0 as 10, which is large enough
to cover the sensitive normalized factors after extensive tests.
On the other hand, given a NF 0, the number of meaningful
candidate scales is also changed when setting different β . Gen-
erally, if β is set large, more segmentation scales are produced,
but the difference is not very significant according to Table 2.
After extensive tests, we choose to set β as 0.9 since the evolu-
tion of meaningful candidate scales is smooth enough.
T able 2. Number of Segmentation Scales within Meaningful Range for the
T est Image T1 When Setting Different Initial Normalized Factors (NF0) and
Decreasing Rates (β ) for Local-Mutual B est Region Merging (LMM), and
L ocal B est Region Merging (LBM) Method
NF0
Number of scales
β =0.8 β =0.9 β =0.95
LMM LBM LMM LBM LMM LBM
20 15 16 19 17 19 21
10 14 14 18 18 20 19
6 12 15 17 19 22 22
3 6 4 6 4 6 4
1 4 1 4 1 4 1
The Effectiveness of AISP on Segmentation Accuracy
In this subsection, the effectiveness of AISP on segmentation
accuracy is analyzed to show how the gradually increased
scale parameters could influence the performance of local-ori-
ented region growing. In Table 3, the segmentation accuracy
is similar for LMM whether AISP is applied or not, and it is also
similar when different NF0 is set. This shows that AISP would
not influence the segmentation accuracy significantly for LMM
method. For LBM method, the segmentation accuracy is simi-
lar when AISP is applied with different NF 0. However, AISP
can improve the segmentation accuracy for LBM significantly,
making LBM achieve similar accuracy as LMM. The segmen-
tation results in Figure 6 further show the difference. The
reasons are illustrated in Figure 7.
T able 3. The Accuracy of Segmentations at Similar Scale for Test I mage T1
W hen Setting Different I nitial Normalized Factors (NF 0). NR Represents
the Number of Regions in Segmentation. The Lowest Row Represents the
Segmentation Result without Using AISP
NF0
LMM LBM
NR BCE Dsym ARI NR BCE Dsym ARI
20 281 0.591 0.495 0.463 250 0.582 0.476 0.494
10 289 0.582 0.479 0.470 245 0.575 0.467 0.500
6 275 0.595 0.500 0.462 260 0.593 0.488 0.474
3 273 0.589 0.490 0.467 230 0.587 0.489 0.494
- 276 0.578 0.476 0.485 245 0.646 0.557 0.420
(a)
(b)
Figure 6. The segmentation results of lbm under the control of (a)
aisp or (b)not. The region number of both results is 245.
Figure 7. A sample explanation of how aisp can improve the seg-
mentation accuracy for lbm method. The arrows in the left part
point to the best neighbor for each region.
466 June 2015 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING