IX-B3, 2012
ven image and is
except for at the
he global inhibitor
) grid. It receives
xerts inhibition to
| for Building
OSM raster data,
nds of differential
y expensive. Thus
ork is proposed.
on from DSM, the
ator estimates the
el location. Given
o establish lateral
Fig.(1) shows the
ent for building
ls i corresponding
ated state. Then
1 the eight cells k
imilarity measure.
,
keNG ©
ce between pixel i
lue of the pixels?
laaximum value of
r block to address
potential for each
Xd it is hard for
by segmentation,
s to participate in
Matrix (GLCM)
Ire, is proposed to
and locate major
ws size is used for
' is represented in
at measures the
chosen to weight
z to their distance
ich are close to |
l.
(6)
j is the number of
tion, the global
tablishes a single
pendently of their
inhibitor plays à
ferent groups. Yet
nce. In this paper,
p between DSM
termination of W.-
For DSM complexity, GLCM contrast attributes are applied to
describe whether gray level distribution is centralized or
decentralized, as contrast returns a measure of the intensity of
contrast between a pixel and its neighbor over the whole image.
The measurement of target occurrence is proposed to show the
complexity of a target and background feature distributions as
well. Target occurrence (R) is defined as Eq.(7), which is based
on analysis of edge level percentages within the image(Mario et
al.,2005). And the inhibition weight Wz is calculated in Eq.(8).
R=P, /(MxN) (7
W,=W «(Target Occurrence + Texture Contrast) x Target Occurrence (8)
Then, one leader cell yet to be excited is selected as a self-
excitable cell. The selected cell is put into the excitation state,
the excitable. Cells are selected based on the coupling weights
between the adjacent cells through coupling term" , and the
selected cells are put into the excitation state. Here, ^! is
applied by logarithmic operation, which was presented by Chen
et al.(2000),which is shown in Eq.(9).At the same time, global
inhibitory takes action to inhibit excited oscillators. Thus,
based on Terman-Wang’s oscillator correlation theory,
oscillators for the same objects can be synchronized, while
global inhibitory is used to discriminate different objects
through de-synchronization. These operations of extended
LEGION are repeated until no excitable cells are detected. If no
excitable cell is detected, inhibition processing is performed,
thereby completing the image segmentation of one region.
These operations are repeated until there is no non-excited and
non-inhibited leader cell any more, thereby pinpointing regions
belonging to the same category from an input image and
identifying them as an image segmentation regions.
Si= Y HS )<R, 11080 N MOI (0
keN (il) keN (il)
After segmentation, morphological cleaning procedures,
such as morphological opening and morphological
reconstruction, are applied to the binary building
segmentation images to remove small objects and to retrieve
the building boundaries that are smoothed out as a result of
the opening operation. Then the building boundary of each
region is extracted from the detected building regions, which
was measured by tracing boundary contours in a binary
image mentioned by Ren, et al.(2002). Since buildings are
regulated objects, solution of least squares with
perpendicularity constraints is put forwarded to determine a
regularized rectilinear building boundaries. Firstly, Douglas-
Peucker-Algorithm gets the feature points through reducing
the number of points of the original tracing point set by
recursively eliminating points that fall below the threshold
of a potential remaining line. In Douglas-Peucker method,
the threshold distance & affected the feature point extraction
directly. Thus, threshold distance £ is defined by perimeter
area ratio. Secondly, the determination of perpendicular
direction is executed for regularizing the boundary. We used
cosine value of near neighborhood pair feature points to
judge perpendicular direction point. If the cosine value of
pair points was less than 0.5, then it was considered as
perpendicular direction point. Finally, all boundary points
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
were applied in a least squares solution with
perpendicularity constraints to determine a regularized
rectilinear boundary. Thus, the polygons are divided into
two groups according to the possibility for perpendicularity
of inner angles of consecutive polygons, and then the
adjustment is performed for each group.
3. Experiment and Results
3.1 Data Description
The test area of Vaihingen in Germany was covered by
altogether 10 strips captured with a Leica ALS50 system,
which was provided by DGPF contains Airborne Laser
scanner (ALS) data (Cramer, 2010) . Inside an individual
strip the average point density is 4 points/m2. The
experimental area is A2 region which is characterized by a
few high-rising residential buildings that are surrounded by
trees. Another test area is a suburban area of scenic place in
China. DSM of of A2 data was interpolated from the ALS
point cloud with a grid width of 25 cm using all return
information. Another experimental DSM data was
interpolated with a grid width of 30 cm. They were
generated by using an interpolation method nearest neighbor
(NN) searches method. When interpolating in two
dimensional space, the especial Quadtree is equal to the
general KD-tree. Thus, space-partitioning data structure KD-
tree was applied for NN search. However, there are some
missing points in raw LiDAR data of urban area. There may
be several reasons causing the missing points. According to
the experimental data of A2 region, LiDAR data gap exists
in significant changes in the ground target’s height. The
missing points are found manually. And the missing points
on the ground are interpolated by neighbor ground points.
Fig.(2) showed the DSM of experimental data.
3.2 Experiment
GLCM homogeneity was applied to distinguish buildings
and tall trees and locate major oscillators in building areas,
which is shown in Fig.(3).Target occurrence and texture
contrast, the parameters of DSM imagery complexity, were
used to define the value of the global inhibitor Wz to
segment pixels into different groups. Tab.(1) showed the
result of DSM complexity measurements and the weight of
inhibition. Thus the extended LEGION scheme was applied
to extract buildings and remove trees from DSM
segmentation in complex urban or suburban areas.
Table. 1 DSM complexity measurements and the weight of
inhibition
Target Texture Estimated
Occurrence | Contrast Real Wz Wz
A2 0.3358 0.1116 38 38.31
Suburban 0.3228 0.8549 31 31.33
Morphological opening and morphological reconstruction
were applied to the binary building images in order to
remove small objects after segmentation. Here we used
square structuring elements in 5 pixels width for A2 area
and 3pixels for suburban area to construct the morphological
operation. The results were shown in Fig.(4).
