NOISE REMOVAL BY THE KARHUNEN LOEVE TRANSFORM
Ram Srinivasan
International Imaging Systems
R&D
1500 Buckeye Drive
Milpitas, CA 95035
Abstract
The Karhunen-Loeve transformation (KLT) or the principal components
analysis has been used to minimize the number of feature sets in
remotely sensed images for better and easier classification. The
KLT compacts most of the energy distributed over several bands in
the data domain to fewer bands in the transform domain. This
process decorrelates the information in the data domain. In this
process it has been observed that the higher frequency components
contain very low variance information, mostly of noise. Mostly the
stripes and banding alone are visible in these components.
The KLT is an orthogonal transformation and its inverse exists.
This paper attempts to remove noise in the KLT domain by dropping
the noisy or low variance components. The remaining components are
used to bring the signal back to the spatial domain. Resultant
images reveal that this process is successful in removing different
kinds of noise (stripe, band noise etc.) that are uncorrelated
between the bands. TM Images have been studied for this spectral
noise removal procedure. Because of the spectral filtering that
this algorithm employs it is expected that this will also
effectively compensate for failed detectors.
roduction
The quality of Landsat MSS and TM images have been analyzed recently
to quite a good extent under the Landsat-D Image Data Quality
Analysis (LIDQA) [1-4] program. One of the major problems reported
and researched by several researchers is the detector and scanning
noise in the images known as striping and banding [3,1]. Several
algorithms have been proposed by researchers to remove the
undesirable noise from the images. The proposed algorithms have
been in the data domain and in the Fourier domain as well. Research
analysts have also been concerned with the problem of failed
detectors [5-6] that results in a null value along a scan line. The
TM sensor has 100 detectors, 16 each in the reflective bands and 4
in the thermal infrared band. In this paper, a method of noise
removal is proposed in the KLT domain. This approach is expected to
yield other benefits in the way of correcting for failed detectors.
The Karhunen Loeve Transformation
In the remote sensing area, the KLT [7-10] is probably better known
263