Overview of Image Processing
81
different polarizations (multi-polarization fusion), different sources (multisource fusion) can be fused to obtain remote sensing products of enhanced
quality.
Fusion of data may be performed at different levels - pixel level, image
level, feature level and decision level (Varshney et al. 1999). The basis of fusion
at any level is again a good quality image registration without which we can
not fully exploit the benefits of fusion. Pixel level fusion requires images to be
registered at sub-pixel accuracy. In this type of fusion, original pixel values take
part in the fusion process thereby retaining the measured physical parameters
about the scene. An example of pixel level fusion is simple image averaging
where the two pixel intensities are averaged. When larger regions of the two
images are used for fusion by applying region based fusion rules, we refer to
it as image level fusion. In the extreme case of image fusion, entire images
are used for fusion. Fusion techniques based on pyramidal decomposition
employ image level fusion (Zhang and Blum 1999). In feature level fusion,
important features are extracted from the two images and then fused so that
features are more clearly visible for visual inspection or for further processing.
In decision level fusion, individual images are first classified individually using
any image classification method, and then a decision rule is applied to generate
a fused classified image that resolves ambiguities in class allocation to increase
classification accuracy.
There exist a number of image fusion methods that may be categorized
into several groups such as transformation based fusion (e.g. intensity-huesaturation transformation (IRS), principal component analysis (PCA) transformation), addition and multiplication fusion (e. g. intensity modulation),
filter fusion, wavelet decomposition fusion, statistical methods, evidential reasoning methods, neural network and knowledge based methods (Bretschneider
and Kao 1999; Simone et al. 2002; Gupta 2003).
In a typical IRS transformation fusion, first the spatial (I) and spectral
(R and S) components are generated from standard RGB image. The I component is then replaced with the high spatial resolution panchromatic image
to generate new RGB image, which is referred to as fused or sharpened image
(Gupta 2003). A novel method based on MRF models to perform image fusion
using hyperspectral data has been described in Chap. 12 of this book.
2.9
Automatic Target Recognition
Automatic target recognition (ATR) consists of detection and identification of
objects of interest in the image by processing algorithms without human intervention. In general, ATR is very difficult. A related and supposedly less difficult
task is automatic target cuing where the algorithm indicates areas or objects
of interest to a human operator. The targets themselves might be actual targets
such as missile batteries in a military setting or areas such as vegetation plots.
The key contribution of hyperspectral imaging to ATR stems from the fact
that target-background separation is seen in the spectral components. It is
81
different polarizations (multi-polarization fusion), different sources (multisource fusion) can be fused to obtain remote sensing products of enhanced
quality.
Fusion of data may be performed at different levels - pixel level, image
level, feature level and decision level (Varshney et al. 1999). The basis of fusion
at any level is again a good quality image registration without which we can
not fully exploit the benefits of fusion. Pixel level fusion requires images to be
registered at sub-pixel accuracy. In this type of fusion, original pixel values take
part in the fusion process thereby retaining the measured physical parameters
about the scene. An example of pixel level fusion is simple image averaging
where the two pixel intensities are averaged. When larger regions of the two
images are used for fusion by applying region based fusion rules, we refer to
it as image level fusion. In the extreme case of image fusion, entire images
are used for fusion. Fusion techniques based on pyramidal decomposition
employ image level fusion (Zhang and Blum 1999). In feature level fusion,
important features are extracted from the two images and then fused so that
features are more clearly visible for visual inspection or for further processing.
In decision level fusion, individual images are first classified individually using
any image classification method, and then a decision rule is applied to generate
a fused classified image that resolves ambiguities in class allocation to increase
classification accuracy.
There exist a number of image fusion methods that may be categorized
into several groups such as transformation based fusion (e.g. intensity-huesaturation transformation (IRS), principal component analysis (PCA) transformation), addition and multiplication fusion (e. g. intensity modulation),
filter fusion, wavelet decomposition fusion, statistical methods, evidential reasoning methods, neural network and knowledge based methods (Bretschneider
and Kao 1999; Simone et al. 2002; Gupta 2003).
In a typical IRS transformation fusion, first the spatial (I) and spectral
(R and S) components are generated from standard RGB image. The I component is then replaced with the high spatial resolution panchromatic image
to generate new RGB image, which is referred to as fused or sharpened image
(Gupta 2003). A novel method based on MRF models to perform image fusion
using hyperspectral data has been described in Chap. 12 of this book.
2.9
Automatic Target Recognition
Automatic target recognition (ATR) consists of detection and identification of
objects of interest in the image by processing algorithms without human intervention. In general, ATR is very difficult. A related and supposedly less difficult
task is automatic target cuing where the algorithm indicates areas or objects
of interest to a human operator. The targets themselves might be actual targets
such as missile batteries in a military setting or areas such as vegetation plots.
The key contribution of hyperspectral imaging to ATR stems from the fact
that target-background separation is seen in the spectral components. It is