sed 
| of 
in- 
(2) 
of 
the 
ess 
ble 
cal 
res 
le. 
cu- 
ra 
ca- 
ble 
ent 
ing 
ca- 
ch. 
in 
sed 
ly 
on, 
ful 
lan 
1es 
not 
ion 
ind 
del 
IX 
/as 
ge 
all 
roads, different kinds of vegetation, and a water area 
(small pond). For the scale space approach, 512 by 
512, 1024 by 1024, and 4096 by 4096 images were 
used. 
4.2 Texture Energy Features 
All fifteen micro-texture images were created in 
coarse level. Then, each image was evaluated to 
select suitable filters. In fact, computational methods 
were not involved to select filters. Based on visual 
evaluation of the micro- and macro-texture features, 
ESES, E5S5, SSES, and S5SS filters were chosen. 
Any micro-filters which did not provide clear texture 
patterns were not used in further levels. Selected 
filters were used through the next two levels. Macro- 
texture features were obtained with different macro- 
window sizes for each level. Macro-filter sizes select- 
ed for each level were 5, 15, and 31 for 512 by 512, 
1024 by 1024, and 4096 by 4096 images, respectively. 
The micro-filters selected in this study provided a 
similar pattern of texture energy features. The filters 
seems to detect horizontal, vertical and diagonal 
patterns of texture. It is obvious that micro-texture 
patterns will disappear by use of larger macro-window 
size. However, more homogeneous macro-textures 
will appear. Grouping of the micro-textures provided 
macro-texture. 
4.3 Integration of Texture Feature Plates 
and Classification 
Each texture feature plate is regarded as a spectral 
band to apply multispectral analysis. The texture 
feature plates are combined to one image file with 
BIL (band interleaved by line) format. Iterative self- 
organizing data analysis technique (ISODATA) in 
ERDAS was used for classification. The advantage of 
ISODATA is that the algorithm represents a fairly 
comprehensive set of additional heuristic procedures 
which have been incorporated into an interactive 
scheme (Tou and Gonzalez, 1974). Classification was 
performed through all three resolution levels. Figures 
3, 4, and 5 are the results of classification for each 
level. The boundaries of both original and classified 
texture images were detected by using a Sobel edge 
operator (Figure 2 and 6). 
The classification results were improved from coarse 
level to fine level. Classification result of fine level 
renders original image. However, the feature 
boundaries were not preserved due to the relatively 
large macro-window size. More boundaries were 
obtained from texture classification by comparing to 
the boundaries of the original image, especially in 
residential and vegetation areas. This result is possi- 
ble, because in those areas different texture patterns 
197 
are mixed. The result still shows lots of micro- 
textures which are reasonable to be grouped into a 
homogeneous area. 
Optimal size of macro-window depends on scale, 
resolution, and objects in the image. To find optimal 
size of the window size is not easy. In addition, other 
very crucial factor for texture analysis is the 
classification method. 
5. CONCLUSIONS 
Laws' texture energy transform provides information 
about texture patterns of the surface. His approach to 
detect micro-texture and then group into a macro- 
texture feature is very realistic and a proper approach. 
However, to determine the fixed macro-window size 
for entire image is a difficult task. It is not easy to 
develop the dynamic size of the window, ie., the 
window size varies depending on the objects in the 
image. 
So far, many of the texture analysis methods do not 
succeed for natural scene imagery. Most of authors 
have used synthetic image or geometrical composite 
(or mosaic) of natural texture image patches to devel- 
op and evaluate texture operators. However, these 
kinds of imagery do not provide enough texture 
properties of natural scene. 
Since color imagery contains more information than a 
monochrome one, to use color image is one way to 
improve the texture analysis system. 
Finally, the 3D object space approach of texture 
analysis is probably a more interesting and more 
powerful solution. 
ACKNOWLEDGMENTS 
Funding for this research was provided in part by the 
NASA Center for Commercial Development of Space 
Component of the Center for Mapping at The Ohio 
State University. 
REFERENCES 
Ballard, D. and C. Brown, 1982. Computer Vision. 
Prentice-Hall Inc., Englewood Cliffs, N.J., pp. 166- 
194. 
Laws, K., 1980. Textured image segmentation. Ph.D. 
dissertation, Dept. of Electrical Engineering, Universi- 
ty of Southern California. 
Gong, X. and N. Huang, 1988. Texture segmentation 
using iterative estimate of energy states. Oth Int. 
Conference on Pattern Recognition, Rome-Italy, 
EDEN 
| 
| 
| 
|