82
2: Raghuveer M. Rao, Manoj K. Arora
almost impossible for a target and the background to offer identical reflectance
in each spectral band. Whereas natural backgrounds typically offer highly
correlated images in narrowly spaced spectral bands (Yu et al. 1997), humanmade objects are less correlated.
ATR may generally be seen as a two-stage operation: i) Detection of anomalies and ii) Recognition. In the detection stage, pixels composing different
targets or a single target are identified. This requires classifying each pixel as
target or non-target using classification approaches described earlier. Thus the
detection stage requires operating on the entire image. On the other hand, the
recognition stage is confined to processing only those regions that have been
classified as target pixels. Recognition also involves prior knowledge of targets.
In many instances, the latter can also be cast as a classification problem. While
there is this broad categorization of ATR into two stages, actual application
might involve a hierarchy of classification stages. For example, one might go
through classifying objects first as human-made or natural (the detection of
potential targets), then classify the human-made objects into buildings, roads
and vehicles, and finally classify vehicles as tanks, trucks and so on.
The early approaches to the ATR problem in images consisted roughly of
thresholding single sensor imagery, doing a shape analysis on the objects of
interest and then comparing the extracted shape information to stored templates. It is easy to appreciate the complexity of the problem, given the many
views, a single 3D object can provide as a function of its position relative to the
sensor with factors such as scaling and rotation thrown in. Furthermore, one
has to operate in poor signal to noise ratio (SNR) and contrast in highly cluttered backgrounds making traditional approaches such as matched filtering
virtually useless. Thus, we may conclude that approaches begun in the early
1980s have reached their limits.
On the other hand, multiband imagery offers many advantages, the chief one,
as already mentioned being the spectral separation that is seen between humanmade and natural objects. An important theme of current investigations is to
collect spectral signatures of targets in different bands. Classifier schemes
are then trained with target and background data for discriminating target
pixels across bands against background. The performance of hyperspectral
or multispectral ATR is yet to reach acceptable levels in field conditions. An
Independent Component Analysis (ICA) based feature extraction approach,
as discussed in Chap. 8, may be used for the ATR problem.
2.10
Summary
This chapter has provided an overview of image processing and analysis techniques used with multiband image data. The intent of this chapter is to provide
a ready reference within the book for the basics of image analysis.
2: Raghuveer M. Rao, Manoj K. Arora
almost impossible for a target and the background to offer identical reflectance
in each spectral band. Whereas natural backgrounds typically offer highly
correlated images in narrowly spaced spectral bands (Yu et al. 1997), humanmade objects are less correlated.
ATR may generally be seen as a two-stage operation: i) Detection of anomalies and ii) Recognition. In the detection stage, pixels composing different
targets or a single target are identified. This requires classifying each pixel as
target or non-target using classification approaches described earlier. Thus the
detection stage requires operating on the entire image. On the other hand, the
recognition stage is confined to processing only those regions that have been
classified as target pixels. Recognition also involves prior knowledge of targets.
In many instances, the latter can also be cast as a classification problem. While
there is this broad categorization of ATR into two stages, actual application
might involve a hierarchy of classification stages. For example, one might go
through classifying objects first as human-made or natural (the detection of
potential targets), then classify the human-made objects into buildings, roads
and vehicles, and finally classify vehicles as tanks, trucks and so on.
The early approaches to the ATR problem in images consisted roughly of
thresholding single sensor imagery, doing a shape analysis on the objects of
interest and then comparing the extracted shape information to stored templates. It is easy to appreciate the complexity of the problem, given the many
views, a single 3D object can provide as a function of its position relative to the
sensor with factors such as scaling and rotation thrown in. Furthermore, one
has to operate in poor signal to noise ratio (SNR) and contrast in highly cluttered backgrounds making traditional approaches such as matched filtering
virtually useless. Thus, we may conclude that approaches begun in the early
1980s have reached their limits.
On the other hand, multiband imagery offers many advantages, the chief one,
as already mentioned being the spectral separation that is seen between humanmade and natural objects. An important theme of current investigations is to
collect spectral signatures of targets in different bands. Classifier schemes
are then trained with target and background data for discriminating target
pixels across bands against background. The performance of hyperspectral
or multispectral ATR is yet to reach acceptable levels in field conditions. An
Independent Component Analysis (ICA) based feature extraction approach,
as discussed in Chap. 8, may be used for the ATR problem.
2.10
Summary
This chapter has provided an overview of image processing and analysis techniques used with multiband image data. The intent of this chapter is to provide
a ready reference within the book for the basics of image analysis.