| do not take
all segments
)ne can take
id adjacency
all segments
m will be eli-
n' from its 4-
ax: Im). differs
n t3/npiy (m).
ent m and t3
liminary seg-
ther layers of
into account
> and spatial
vestigated in
Id recognize
nay itis
instant value
nt m. Such a
the original
, and it often
n the gmean -
noothed ver-
stressed that
tion Lab(i,j)
it number to
ment can be
/ values gi j-
segment fea-
higher layers
and object
od the LGN
uter (486 PC
A number of
images was
lds t4(l) and
= 1, mex 1)
5 gave best
more than
f(m,i,j) used
ith less than
"he following
er values.
e image was
). The result
ation of small
There is no
c. Essential
ined. Fig.1d
3 segmented
without (9),
(10)). The result (fig.2b) has 1236 segments, the biggest
segment with 2056 pixels. After the elimination of small
segments we have only 744 segments (fig.2c) which are
sufficient for describing this scene with many small details.
The 10 biggest segments are shown in fig.2d. One can see
the complicated structure of some of these segments.
The image of fig.3a (a fjord region) was segmented with
t; 70.6 and t576. Without (9) and (10), after the elimination
of small segments one obtains 1702 segments (fig.3b).
Taking into account (9) and (10) only 1408 segments are
left (fig.3c). These segments describe most of the scene
adequately, but the segmentation of the clouds (lower left)
is not sufficient. This shows that a modification of (10)
must be investigated. The 3 biggest segments (fig.3d)
once again show a complicated, fuzzy’ structure which is
best described by the PAG.
The forest image (fig.4a) which was segmented with
1170.65, t5-7 and taking into account (9) and (10) has
2301 segments (fig.4b). This number reduces to 1510 after
eliminating small segments (fig.4c). Many of the small seg-
ments which are retained represent the forest textures,
and hopefully can be used as texture elements for texture
segmentation in higher layers of the network. The 10 big-
gest segments (fig.4d) mainly represent parts of the back-
ground.
References
Haralick, R.M., Shapiro, L.G., 1985. Image Segmentation
Techniques. CVGIP 29, pp. 100 -132.
Jahn, H., 1986. Eine Methode zur Clusterbildung in metri-
schen Räumen. Bild & Ton 39, pp. 362 - 370.
Jahn, H., 1996. Image Segmentation with a Layered
Graph Network. SPIE Proceedings, Vol. 2662, 1996 (in
press)
Levine, M.D., 1985. Vision in Man and Machine. Mc Graw-
Hill, New York.
Pavlidis, T., 1977. Structural Pattern Recognition. Sprin-
ger-Verlag, Berlin.
Uhr, L., 1980. Psychological Motivation and Underlying
Concepts. In: Tanimoto, S., Klinger, A. (Eds.). Structured
Computer Vision. Academic Press, New York.
Figure 1: Mars image
Figure 3: Fjord region
Figure 4: Forest image
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
A
ERES
ERES
I
