292
12: Teerasit Kasetkasem, Pramod K. Varshney
The results of changed sites for individual spectra are displayed in Fig. 12.7a,b,
12.7c,d and 12.7e,f, respectively. Figure 12.7a,c and e are determined by the
MRF based ICD algorithm while Fig. 12.7b,d and f result from image differencing. By carefully comparing results within each color spectrum, we conclude
that our ICD algorithm detects changes that are more connected than those
found using image differencing.
12.4
Image Fusion using an MRF model
The idea of combining information from different sensors both of the same
type and of different types to either extract or emphasize certain features
has been intensively investigated for the past two decades (Varshney 1997).
However, most of the investigations have been limited to radar, sonar and
signal detection applications (Chair and Varshney 1986). Image fusion has
gained quite a bit of popularity among researchers and practitioners involved
in image processing and data fusion. The increasing use of multiple imaging
sensors in the fields of remote sensing (Daniel and Willsky 1997), medical
imaging (Hill et al. 1994) and automated machine vision (Reed and Hurchinson
1996) has motivated numerous researchers to consider the idea of fusing image
data. Consequently, image fusion techniques have emerged for combining
multisensor and/or multiresolution images to form a single composite image.
The final composite image is expected to provide more complete information
content or better quality than the individual source image.
The simplest image fusion algorithm is the image averaging method (Petrovic and Xydeas 2000) in which a fused image is a pixel-by-pixel average of
two or more raw images. This technique is very robust to image noise when
raw images are taken from the same type of sensor under identical environments (Petrovic and Xydeas 2000). However, if images are taken from different
types of sensors, the image averaging method may result in the loss of contrast information since the bright region in an image obtained by one type of
sensor may correspond to a dark region in other images taken from different types of sensors. Furthermore, the image averaging method only utilizes
information contained within pixels even though images are known to have
high correlations among neighboring pixels. To utilize this information, Toet
(1990) developed a hierarchical image fusion algorithm based on the pyramid
transformation in which a source image is decomposed successively into a set
of component patterns. Each component pattern corresponds to the representation of a source image at different levels of coarsenesses. Two or more
images are fused through component patterns by some feature selection/fusion
procedures to obtain the composite pyramid. The feature selection procedure
must be designed in such a manner that it enhances the information of interest while removing irrelevant information from the fused image. Then, the
fused image is regenerated back through the composite pyramid. Burt and
Kolczynski (1993) have developed a similar algorithm with gradient pyramid
transformation. Uner et al. (1997) have developed an image fusion algorithm
12: Teerasit Kasetkasem, Pramod K. Varshney
The results of changed sites for individual spectra are displayed in Fig. 12.7a,b,
12.7c,d and 12.7e,f, respectively. Figure 12.7a,c and e are determined by the
MRF based ICD algorithm while Fig. 12.7b,d and f result from image differencing. By carefully comparing results within each color spectrum, we conclude
that our ICD algorithm detects changes that are more connected than those
found using image differencing.
12.4
Image Fusion using an MRF model
The idea of combining information from different sensors both of the same
type and of different types to either extract or emphasize certain features
has been intensively investigated for the past two decades (Varshney 1997).
However, most of the investigations have been limited to radar, sonar and
signal detection applications (Chair and Varshney 1986). Image fusion has
gained quite a bit of popularity among researchers and practitioners involved
in image processing and data fusion. The increasing use of multiple imaging
sensors in the fields of remote sensing (Daniel and Willsky 1997), medical
imaging (Hill et al. 1994) and automated machine vision (Reed and Hurchinson
1996) has motivated numerous researchers to consider the idea of fusing image
data. Consequently, image fusion techniques have emerged for combining
multisensor and/or multiresolution images to form a single composite image.
The final composite image is expected to provide more complete information
content or better quality than the individual source image.
The simplest image fusion algorithm is the image averaging method (Petrovic and Xydeas 2000) in which a fused image is a pixel-by-pixel average of
two or more raw images. This technique is very robust to image noise when
raw images are taken from the same type of sensor under identical environments (Petrovic and Xydeas 2000). However, if images are taken from different
types of sensors, the image averaging method may result in the loss of contrast information since the bright region in an image obtained by one type of
sensor may correspond to a dark region in other images taken from different types of sensors. Furthermore, the image averaging method only utilizes
information contained within pixels even though images are known to have
high correlations among neighboring pixels. To utilize this information, Toet
(1990) developed a hierarchical image fusion algorithm based on the pyramid
transformation in which a source image is decomposed successively into a set
of component patterns. Each component pattern corresponds to the representation of a source image at different levels of coarsenesses. Two or more
images are fused through component patterns by some feature selection/fusion
procedures to obtain the composite pyramid. The feature selection procedure
must be designed in such a manner that it enhances the information of interest while removing irrelevant information from the fused image. Then, the
fused image is regenerated back through the composite pyramid. Burt and
Kolczynski (1993) have developed a similar algorithm with gradient pyramid
transformation. Uner et al. (1997) have developed an image fusion algorithm