ul 2004
\pproximation
age (change)
Horizontal detail
image
Vertical detail
image
Diagonal detail
image
sample
wo-siep
spatial
vith the
ions are
(ep are
versely
multi-
spatial
NANT
object-
] image
| layout
lasses
largely
sector,
ze filter
ision
cution |
tA Ae
RGB HS |
CTrol
CIToR
Tu RED |
IHS—RGB |
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
Raw materials are integrated and managed on CManage class
and CFusion class is responsible for image fusion processing.
Each algorithm for image fusion processing is configured to
components.
4.1 Development of Relevant Components
Components developed related to this study were classified
into total five categories according to component functions.
Classified categories are data input, image filtering, image
analysis, image fusion and output component. The details are
explained below.
1 - Data Input Component
Image Loading Library(Tif, Jpg, Raw):
CSatelmage Class Allocation
2 - Filtering Component
Image Improvement Component : CFltImage Class
3 - Image Analysis Component
Analysis of Fusion Images : Normalization,
Equalization, Histogram Component
4 - Image Fusion Component
Integrated Fusion Component Management
Class
Analysis of Fusion Images: FusionImage Class
Optimization of Memory Use : CRetlmage Class
5 - Output component
Output Component : CFusionlmage
CFusion
4.2 Developed Program Format
Program window in Figure 5 shows multi-band image input
window. Overall program menus comprise of Enter and Save
File, Image, Layer, Select, Fusion, Tools, Windows and Helps.
Submenu on the left side consists of Fusion Method, Pass Step,
Rough Images on Pass and Display Window of Image Size.
© CIO
NER RE ov
Fusion Method |
Pass Depth
Edge Process
— =
WH APass 18 p
Figure5 Multi-band display of fusion program.
5. APPLICATION TEST AND ANALYSIS
5.1 Application Test
This study used multi-band images from LANDSAT taken
around the Hangang river in Seoul on April 4, 1995 and images
from KOMSAT-1 taken on the same area at 9 minutes to 2 on
March 1, 2000. The same points on each image were manually
selected and finally selected using Brightness Correlation.
Geometric transformation formula between two images was
defined using Affine Transformation and images were
rearranged by bilinear interpolation method. Next, two images
were created for image fusion and IHS and Wavelet fusion
methods were applied to those two images. HPF fusion images
had relatively inferior resolutions so that they were excluded
from analysis.
5.1.1 IHS Method
Low resolution images from LANDSAT were prepared by
band and were transformed to Intensity(l), Hue(H) and
Saturation(S). In Figure 6, 1, 2 and 3 means the upper limit of I,
H and S, respectively. images on the upper limit 4 are fusion
images. Dominant appearance of yellow was phenomenon
caused by fusion because the angle of Hue was expressed only
to | in an angle unit.
Intensity images were reversely transformed by RGB and the
results were shown in Figure 8. R, G and B mean |, 3 and 4
quadrant, respectively. These images were fused using
intensifies of high resolution images. At this point, images were
replaced by pixel values corresponding to each image size.
Consequently, for images in 1:4 size, the intensify of one
pixel of low resolution image was replaced by four pixel values.
On the result image on the left side, slightly light colors were
appeared on mountain areas and for areas with higher
brightness, blue or sky blue color was shown. In terms of
calculation, we could get values exceeding 0-255. These values
were obtained because relevant values were lost during linear
stretching for images to adjust these values. The final fusion
images were presented on the left side of Figure 6.
Le me
Figure6 RGB reverse transformed image and fused image.
5.1.2 Wavelet Method
Wavelet transformation was applied to low resolution
images from LANDSAT as well as high resolution images from
KOMPSAT. As shown in Figure 2, these processes passed
through low pass filter and high pass filter. The images in the
size of 512*512 were reduced to 256*256—128*128—64*64
—32*32. Therefore, we had limit that image sizes should be 2”.
In general as Wavelet transformation was processed in
multiple steps, spectral resolution of low resolution images and
spatial resolution of high resolution images could be fused
much better. In accordance with analysis using a developed
system, when image transformation exceeded appropriate
image compression rate in Wavelet transformation, images
were transformed and so influenced on image fusion. Images in
Figure 7 described that image resolutions were increased as
result passes of each band according to the number of each pass
were repeated. It was found that four Wavelet transformation
steps created the best image fusion in accordance with system
analysis. From fifth step, images were transformed so that
image fusion was not properly carried out.
This wavelet transformation was applied from step 1 to step
4 as passing through low resolution image and high resolution
image by step. When rough images and precise images were