If it is set larger than the sensitive normalized factors, it can
make sure to cover the meaningful scales.
β
determines the
smoothness and continuity of the increased scale parameters.
The experiments show that the increment of scale parameters
is not very sensitive to
β
, as shown in Table 2. Hence, it is not
difficult to set the two parameters according to the analysis.
The effectiveness of
AISP
on segmentation accuracy for lo-
cal-oriented region growing is also analyzed to reveal how the
gradually increased scale parameters can influence segmenta-
tion accuracy. The results show that
AISP
has no significant
influence on the accuracy for
LMM
, but it can enhance the op-
timization ability of
LBM
by changing the growing region and
allowing region pairs with large similarity to be merged first.
The analysis presents a manner to improve the performance
for local-oriented region growing, which is that we should
not keep growing a region until it satisfy the stopping rule but
change the growing regions after several iterations to merge
other region pair with greater similarity first.
AISP
controls to produce nested multi-scale segmentations,
which serve as candidates of meaningful segmentation scales.
In the following step, we should select meaningful scales for
given applications. To accelerate scale selection, the segment
tree is used to represent the nested multi-scale segments,
where the nodes at different levels are hierarchically linked. It
is very fast to produce multi-scale segmentations by cutting the
segment tree at different levels. In this study, the segmentation
scales for evaluation and shown are selected by visual analy-
sis. Owing to the efficiency of cutting segment tree, we suggest
to setting the target scale parameter large to build the segment
tree. Then, the segment tree records segmentation scales coarse
enough to cover meaningful segmentation scales and we can
select meaningful segmentation scales by cutting the segment
tree. In the future, we would pay attention to choosing mean-
ingful scale(s) for a given application automatically. Then,
there is a need to build the correspondence between the scale
parameter and the semantic meaning of objects.
Conclusions
The adaptively increased scale parameter (
AISP
) strategy has
been presented for multi-scale segmentation of high-spatial
resolution remote sensing images.
AISP
is used to control
region-growing procedure to produce nested multi-scale
segmentations. In this study, we focused on dynamically
determining a set of gradually increased scale parameters in
AISP
, which are adaptive to specific images so that they can
capture the variety among different regions. The normalized
factor was introduced to calculate scale parameters, which
contributes to form continuous upscaling mechanism to cover
meaningful scales. The experiments proved the effectiveness
of
AISP
on controlling multi-scale segmentation. Especially,
the influence of gradually increased scale parameters on
segmentation accuracy was analyzed in experiments, indicat-
ing the importance of changing growing regions for improving
segmentation performance of local-oriented region growing.
The comparison with
MRS
method showed that
LMM
+
AISP
can achieve slightly better performance.
Acknowledgments
This work was supported by the Science and Technology Pro-
gram of Zhejiang Province (Grant No. 2014F50022) and the
Qinglan Project of Jiangsu Province (Grant No. 201423). The
authors thank the editors and anonymous reviewers for their
suggestions improving the paper.
References
Baatz, M., and A. Sch pe, Multiresolution segmentation - An
optimization approach for high quality multi-scale image
segmentation, Angewandte Geographische Informations-
Verarbeitung XII (J. Strobl, T. Blaschke, and G. Griesebner,
editors) Wichmann Verlag, Karlsruhe, pp. 12–23, 2000.
Beaulieu, J.M., and M. Goldberg, 1989. Hierarchy in picture
segmentation: A stepwise optimization approach,
IEEE
Transactions on Pattern Analysis and Machine Intelligence
,
11(2):150–163.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 9. Multi-scale segmentation results of T3 (top row) and T4 (bottom row) using LMM+AISP. The number of regions is 1174, 335,
204, 61 in (a) through (d), and 226, 166, 83, 33 in (e) through (h).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
469
