The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
In a word, the MHR not only considers colour information, but
also shape information which could reduce the disturbance of
noise, and debase fragmentized degree of object boundary, and
get more regular objects.
2.3 Multi-scale Segmentation based on SRM and MHR
On the basis of analysing the SRM and MHR algorithm, a new
multi-scale segmentation algorithm is presented, where the
SRM is improved and applied in HR imagery for getting initial
fine segmentation results, and the MHR is used for merging two
adjacent regions from the initial segmentation results.
The improvements of the SRM algorithm itself are as follows:
(1) Make full use of all bands information of remote sensing
imagery. The original algorithm is only suitable for grey and
colour image, we improved and applied it in remote sensing
imagery with many bands.
(2) Define scale parameter S and set up the relationship
between 5 and the independent random variables Q. Since the
SRM has the ability of multi-scale segmentation, S is defined to
satisfy the direct relationship between the imagery scale and the
object size. Namely, the bigger the scale is, the bigger the
object size is, the more the object numbers are.
(3) Define merging predicate more strictly. Since
y ]b 2 (R) + b 2 (R') <b(R) + b(R r ) , the strict merging predicate is
therefore:
/»(/?,*') = { true '
[ false.
if Vae{B\,B2,...,Bn},\Ra-Ra |< b(R) +b(R')i^)
otherwise
Where,
b(R) = g,
10000
2S\R\
, S is scale parameter,
R a denotes the observed average for channel a in region
stands for the set of regions with R pixels. B\,B2,...,Bn
are channels of imagery.
Therefore, the flowchart of the multi-scale segmentation
algorithm based on SRM and MHR is shown in figure 1 which
comprises two main progresses: the initial segmentation
progress by the improved SRM algorithm, and the merging
progress by the MHR algorithm.
tpixei based segmentation} t tobjeef- merging}
i
I
I
I
I
I
I
I
I
«
1
I
-J
Figure 1. The flowchart of the Multi-scale segmentation
based on SRM and MHR.
(2) Ascertain the merge predicate shown in formula (2) which is
relative with pair-pixels, and make sure the position of up level
nodes the pairs belong to.
(3) Judge whether the seeds of pair-pixels are at the same
position, and whether they satisfy the merge predicate. If their
positions are not identical and they satisfy the merge
predicate(.S<777/), then the pair-pixels are merged, meanwhile,
the area is updated with the sum of the pair-pixels.
(4) Repeat step 2-3 until all the pair-pixels are segmented by the
approach. Then an initial segmentation result which is based on
pixel-based segmentation is realized.
As to the multi-scale segmentation, there are four steps in the
merging progress:
(1) Object polygons are generated by vectorization algorithm.
The key steps are the search of boundary, the generation of
topology structure, and the remove of redundant points. Then
the information such as topology structure, pixel numbers,
mean, deviation and boundary length are stored in a vector file
and an attribute file.
(2) Set the parameters of MHR, such as
w , w , w , w L > Th2. And then compute
W color ’ W shape W compl ™smooth ’ r
heterogeneity value h of neighbour polygon according to
formula (3).
(3) Judge whether h satisfy MHR, if h<Th2, the adjacent
smaller objects are merged into other bigger ones, meanwhile,
the average size, deviation and mean of all the object regions
will be calculated.
(4) Repeat step 2-3 to accomplish multi-scale segmentation.
This relationship between each level is shown in figure 2.
Figure 2. Four-scale hierarchical network of image objects
Scale 1 stands for initial segmentation level based on pixel-
based segmentation, scale2, scale3 and scale4 stand for multi
scale segmentation level based on object-based merging. To
guarantee a definite hierarchy over the spatial shape of all
objects the segmentation procedures follow the following rules
( Benz, 2004):
(1) Object borders must follow borders of objects on the
lower scale.
(2) Segmentation is constrained by the border of the object
on the upper level.
(3) The correction of object shape based on merging sub
objects is possible.
There are four steps in the initial segmentation progress:
(1) Set the sort function shown in formula (1), and then sort the
pair-pixels according to the size of the function.
1259