and grouping. To develop a computational texture
analysis system is not an easy task due to the great
complicatedness of properties of texture.
Texture has following characteristics;
* Texture is shift, orientation, moment, contrast, and
illumination invariant.
* Human texture perception tends to be sensitive first-
and second-order statistics, and does not respond to
higher than second-order. Discriminable textures can
be generated having a common mean, variance, and
auto-correlation function. Thus, second-order mo-
ments are sufficient measures of texture.
* Texture is hierarchical, i.e., it corresponds to
different resolutions and then global unitary
impression is offered to the observer. Global features
characterize the whole texture rather than texels.
The above characteristics are helpful guidelines in
designing texture analysis system. The other
important characteristic is that texture is both
stochastic and deterministic, therefore, texture
analysis methods are categorized by two major
approaches; statistical and structural approaches.
3. TEXTURE MEASUREMENT
Texture energy transform developed by Laws is a
class of spatial-statistical approach. The characteristic
of this method is more matched to intuition about
texture features, i.e., similar to human visual process-
ing (Laws, 1980; Ballard and Brown, 1982). This
method was developed after he investigated and
evaluated several existing methods including statisti-
cal, structural, co-occurrence, spatial frequency, and
auto-correlation approaches.
3.1 Texture Energy
The original image or a patch of the original image (f)
is convolved with micro-texture filters (hy) to create
micro-texture features (f'k);
FD - fap (1)
where the micro-texture filters can be formed from
following four one-dimensional vector masks;
n5 - [1° 4 6°>4 11
h5 = [-1 +2 0 2 1]
55 = [FL 60 2 0x1]
R5 [.1.—-4.:6:-4. 1]
A total of sixteen two-dimensional micro-texture
filters can be created. These are L5L5, LSES, L5S5,
L5RS, ESLS, ..., RSR5. However, LSL5 is not used
because the sum of the filter elements is not zero.
In order to obtain macro-texture features (f"k), each of
the micro-texture images (f) is transformed into an
texture energy image by moving macro-texture win-
dow;
w,
rape dez
ij = —
k w? n=v% m=%
f', (n,m) | (2)
where w is size of a macro-window. The micro-
texture feature values are replaced by average of
absolute values in a macro-windows. The size of the
optimal macro-window depends on texture coarseness
or regularity, as well as the quality of the available
micro-features.
Micro-texture filters are designed to measure local
texture properties, while the macro-texture features
measure properties of the texture field as a whole.
The problem is there is no guarantee that any particu-
lar resolution or window size will be optimal for a
given analysis (Laws, 1980).
3.2 Texture Classification
‘Texture segmentation can be performed by classifica-
tion. Most of the classification algorithms are suitable
for multispectral imagery. Since several different
micro-texture filters provide many corresponding
texture feature plates, to use a multispectral classifica-
tion algorithm is a quite reasonable approach.
Classification of imagery is one of the main tasks in
remote sensing. However, the pure texture-based
classification method does not seem to be successfully
developed yet. The purpose of image segmentation,
based on texture information, is to obtain useful
surface information.
Unsupervised classification is more attractive than
supervised classification methods, because sometimes
a priori knowledge about the area of interest is not
available. Furthermore, human operators' intervention
will not be allowed in fully automatic mapping and
surface reconstruction systems.
4. EXPERIMENTAL RESULTS
4.1 Selection of Imagery
Left image (photo scale: 1/3,800) of "Munich" model
(Figure 1), which was digitized with an EIKONIX
camera to a resolution of 4096 by 4096 pixels, was
used to implement our task. The "Munich" image
contains residential areas, major high-ways, small
196
TO:
lai
th
re
bl
