b'r n U C T' U F< (1 L
PI
(:1 F' F' 1::< D (4 C I···!
'r I..·j E
rU
U r:' 'r EX ' 1' U F: E F' E {:I ' 1' U F< E~ 5
.~: X 'r F:': PI C 'r I U 1\1
FROM SPC1 AND eIR PHOTUGRAPHS.
1 .. ' 1' .. C ..
p
q
1-3 CJ >:
(J"
7500
I~)(.~
(;; ..
Enscl·"I(?dE'
f··,iE!:tl"'f;:?r l.i::tnds
111/ il·
A structural approach to the quantification of texture is
Texture is treated as a state of an area between
pure pattern and pure noise. As topological relations play an
i rn p 0 r" t:. E:\ 1"'1 t: r' ul
E~
:i.
C:i.U r"
I"i
In ('!::' t:. h
r' (.: ., 1 i::\ t :i. \/ :;.,~:I. Y :i. n ~::~, c:·! ('1 ::::. it:. :i. v \::.~ t
D
'1':.1'' ) (.:":, c:I E~ ~... :i. v E·: c:!
C)
d
C
I"', to: ' i"', (J F::: ~::i, i
1''',
1'- ~.~\ cJ:i.
+E: E:'t t
Cj
rn E,! t
u I'" i:?~ S; i::( 1'- (::..:'
1'" \;
q E~ D rn E,' t:. I"" y •
0 t..·
I !\IT'HUDUCT IOl\!
The classification of areas based on textural features is
usually approached using a statistical method such as auto
correlation,
tables.
local variance or value transition frequency
As these methods are inherently based on a noise model
we decided to approach the extraction of textural features from
the structural pattern recognition side.
We start with the
analysis of structured patterns which can then become less
structured by the addition of noise or less discernable because
of low resolution.
In order to evaluate the performance of our method we
i::\ppl:i.f:!!d
it, t:Ci
elements)
(:j ~::i,
Ci /,,'1
c::'
lDVJ
1'''('':~l:;i,Cilu.t:i.Cln
ElPCJ'T c:Ii::\t:.a
~5CE~nf.~
(l(Zii/i ,,:,\r"le:! 2(Z)m
and to vegetation canopies in high resolution
D
+
t
;'''1
(.:.:~
I'" (-;::' ':::i, C::.' <:). 1'- C
!'!
i:':', :L rf'i ;.,;
:i.
E
t
D
d (0: :~;:i. (] n
<',:\ n
up t :i. c a. 1
(0.25m)
t
E!;';
t, u r' (;2
channel
,the similation of a set of optical spatial filters on
thE'! b {::\ :~:. :i.
~::;
0+
thE'
is reported first
d ~.:'t t: (::t
n
j"'!i:':\c:! E:".rnC':'tl''' d··..
<:':tr'l d
PI''':i. n c:: i p a].
c: [:imp on ('21''', t
t.l~ dn ;5+ (:ir rns
.The data used are the high resolution elR
''I'' I'' , E:: :::;:L rn u 1 <':':'. t F,: d
C) P (,?~, I'" i:':\
t
c) 1'''
:;;i, i ;:':. (:~!
:i. ;::i, /1, ;<
111 ... 406
4
F:? 1 (~? rn F:~ n
t
:5 •
We review the approach of correspondence analysis, next we
report results on a limited number of test areas and test data.
CLASSIFICATION BASED ON TEXTURAL FEATURES.
Texture is a feature of an extended area . The area shows a
variation in observable attributes (features) which are
characteristic for the class the area belongs to.
Correspondence analyses [Mulder,Radwan ISPRS88J is applied
to subarea (segment) patterns. Subarea spectral or structural
features define local feature vectors. Within a limiting search
neighbourhood other subareas with maximum correspondence and
minimum distance in feature space are found. The network of
corresponding subareas has topologocal and distance properties
which are typical for the pattern or texture under observation.
Completely random data will have random, nonorganised network
features. Stuctured data such as generated by images of modern
suburbs will generate a network reflecting the structure of
blocks of housing and the surrounding roads.
Areas of constant texture are outlined by merging segments
of the previous level (of merging) based on correspondance
analysis applied to these segments.
The problem of classification of areas on the basis of
pattern which may degenerate to almost pure noise can be
approached from two extreme sides. One is to assume the data is
generated by a noise source with peculIar parameters. The other
model is that the objects in the data are basically forming a
structured pattern but there is insufficient resolution or
noise which distorts or hides the pattern.
In order to describe pattern in an area, structure units
have to be defined. Structure units follow from low order image
segmentation (over-segmentation, fine grained segmentation).
Each small segment has a list of attributes (segment features).
Correspondence analysis is applied first to the adjoining
neIghbours. As often patterns are generated by objects against
a background corresponence analysis has also to be applied to
111-407
obj
t7: C
t: 5
It'J
hi ch
i::\ r~
e a
II
dis t: (':i\ n c E'
II
a p ii:\ r" t .
In some applications the distinction of objects and
background must be solved before the correspondence analysis
1S
applied. A reason for this could be that the network pattern of
corresponding segments in the foreground is not at all related
to the network pattern of the background.
In any structural approach tHe use of knowledge is critical.
Therefore the classification process must be preceeded by some
sort of supervised training and explicit formulation of pattern
modt·:d s.
(JUI'" 1'''(':?SE'('::\I~ch on DElvl gf.0nE::I'" (':7lt i 01"1 +r'om
r~:;tE·!r"(:7:o
bPUT d(::'.t (:;i.
II
II
indicates that a measure of self-correspondence (texture)
must
be available at the time when cross-correspondence is
evaluated. This is equivalent to the application of cross
correspondence analysis to le+t and right images which have
already been segmented to the level including textural
f
(-?~ i:i\
t u I'" (:;.~ ~:~ ..
In the poster session we will show the result of a number of
(~0H
P E'! I'" i rnf2n t. s
In {Gorte, Mulder,
A)
1987] examples are given on the
discrimination of artificial teHture and texture occuring in
bPOT images using the density of structure-class symbols. This
approach was as successful or slightly better than the approach
as followed by Granlund using the GOP (Context Vision) machine.
A good discrimination was achieved between build-up areas and
other.
This discrimination was not possible on spectral
features only. The addition of the textural feature was
E:f:5l::;E?nti d.l ..
A recent improvement is the availability of the distance
t I'· ii:\ n ~:; +0 I'" in 0 n t. h <:::.. [30 P ( Con t. E' :.~ t \J i !::) :i. 0 n) ii:l. r"l d i::\ +a ~::; t d ('Y!: n ~s i t Y
transform implemented on the GOP
B)
(Gorte,1988) "
Design of an optical texture channel ..
Data :digitised elR,
application :vegetation canopies, tree stand and species
di
SiC''''
i m:i. nat ion"
111-408
types is not well possible. To improve spaceborn R8 systems
Mulder proposed in 1983 to add on optical texture channel for
the discrimination of vegetation canopies and landuse classes.
The design criteria were to have a ground resolution of 1 to 2
m and a texure element matrix of 4 x 4 to 8 x 8 elements. The
optical texture masks would be generated either from Hadamard
matrices or derived from an other set of orthogonal filters
(spatial principal components).
Image 1 shows the original IR band at a resolution of 0.25 m
and 16 Hadamard components (data +127) derived from 4 x 4
Hadamard matrices in pseudo greyscale.
Image 2 shows the 16 principal components of 4 x 4 subimages
(data + 127) represented as greyscale images.
In order to contrast the structural method with the
statistical method ,image 3 shows the local variance applied to
the same 4 x 4 subimages.
The next step is to select a subset of textural features and
use them as input for the segmentation programme. The
segmentation programme produces a segment property list and a
region (segment) adjacency graph which is the required input
for the structural texture analysis procedures.
C)
Texture analysis based on using the segment property list
and the region adjacency graph.
:i. n c:: Iud E:! l::;
~
(;'('J
y ) ·f i
I'" s;; t
..- s can _.. i
Presently the property list
I"l t (':01'" s G! c
t:. ,
(x1,yl,x2,y2)boundil"lg-rectangle
l;:; E~ <;J m (~n t
..- a I'" f:0 ii:\
nr-of-adjacent-regions.
Correspondence analysis is implemented as a search for
adjacent neighbours which have less than a threshold distance
in feature (property) space.
As segment area and number of neighbours are both related to
the density of segments they are used as first candidates for
further segmentation (merging). The average density and average
number of neighbours over the new
(texture) seqments form part
of the new properties of segments at the present level.
Examples will be shown during the poster presentation.
CDI\ICLUB I UI\18
111-409
- with high resolution data it is recommendable to preceede
structural texture detection by low order pattern detection
filters. The prospects for an optical texture channel are
encouraging further R&D and more effort in marketting the
concE~pt ..
- using correspondence analysis for directing the high-order
merging of segments works well in the absence of background.
- the interaction of foreground connection nets and background
connection nets has to be studied further ..
- the selection of segment properties (features) and
subsequent decision parameters has to become more knowledge
dri v(-?:!n" (.:it prf!:!sE:'nt (Jood I"'l:~f::;ul t~::~ <:;\Y--E\ mos~tl y obt<::\:i_ nE-:d by manual
-f i n Fi! t. un i r-I lJ
b -y' t: h 0:: I/~ f:0Sif!i!al'" c h E:lY-- ..
II
II
- there are indicat.ions t.hat t.he st.ruct.ural approach is bet.t.er
suited for int.erfacing with knowledge engineering systems t.han
t.he stat.istical approaches to texture feat.ure extract.ion.
[Jew- -t (::i! D .. D .. 1---1..
t-"iu I
ON TOPOGRAPHiCAL OBJECTS
d E-:!"-- 1\1 .. J ..
1 (jiD"?
'!
ffiCX'1 REl'K)'fE SENSING DATA '(
1 9UB
transform on the BOp l1
(oct.ogon)
/, BUILnING
ASSISTANTS J:<UR THE .IDm?ACrrON
SYYnp·)
Imp 1 F:!rni:?i!n t_ i n q
v...J,'LLi No rd.. beA'J
'j
II
C:; ..a:z..
f (::\ ~::; t_ d :i_
E\
~s
t
<::\[""1
C (::£,
, Technical Report TR8802,
ITC/IPL, May 1988.
[iOI,.-tF!:! B .. C3 H ..
ti"'i;:tr1!sfor-rn on
M
1 9E3B
'.I
th(-:~!
BDF"I
II
'!
T(:::!chrlical
II
I rnp].
~?:!mE:'n t_
in q <-,:\
Technical Report TR8803,
II
al(JOi/~ithmll
'.I
II
+ii;\ S t
tY
ITC/IPL, May
cl f!:!ri Si i
Proc.ISPR88B, Kyoto.
Proc.ISPRS88, Kyoto.
II
F~~~pOi/--t
THt38(ZJ,Q,
111 ... 410
ITC/IF'L,
(in
PI"'E~pi:;tri0.tion)
Image 1: IR band ( 0.25 m ) digitized form a eIR aerial
photogragh ( 1:2500 ) and Hadamard components
111 ... 411
Image 2: Spatial principal components ( derived from 4*4
subimages )
.
Image 3:
Local variance ( derived from 4*4 subimages).
111 ... 412