Speckle is the major multiplicative noise in the SAR(Radar) images, Improvement is done by using stochastic distance methods by assuming data as gamma distribution which enhances the images by 78% overall....
1. Despeckling SAR Images using Non-Local means Filter
with Stochastic Distances
BY PANKAJ KUMAR
U.G. BTECH CIVIL ENGG
2. Introduction
SAR Image: These are the images created by using Radio-waves collected using
radar. SAR uses the motion of the SAR antenna over a target region to provide finer
spatial resolution than possible. Spectrum Range is 30 MHz to 30GHz.
SAR image vs Optical image
SAR Image Optical Image
Consists of frequency region higher than
Far Infrared till Very High Frequency
Consists of frequency region in visible
spectrum till NIR
Day/Night images are equivalent Temporal effects change the images
Provides Equivalent images in all weather
conditions
Changes pattern of images in different
weather conditions
Highly sensitive to moisture Less sensitive to moisture
Can’t be seen by naked eye, needs radar
to sense intensity values
Can be seen by naked eye
4. Speckle
Speckle is a granular 'noise' that inherently exists
in and degrades the quality of the active radar ,
Synthetic Aperture Radar.
Speckle is primarily due to the interference of
the returning wave at the transducer aperture.
It is caused by coherent processing of
backscattered signals from multiple distributed
targets.
Several methods and approaches have been
considered to eliminate this noise but due to its
multiplicative nature of the noise it is difficult to
remove it.
5. Methods to remove Speckle consists of three types of filteration techniques:
Multiple-look processing of image
Adaptive filteration
o Non-Local means filter by Alejandro C. Frery & L. Torres
o PPB(Probabilistic Patch-Based) approach by Charles Deledalle
Non adaptive filteration
o Mean and Median Filter
o Wiener Filter by Norbert Wiener
o Frost Filter
Thresholding Based filteration
o Wavelets Thresholding methods
6. Novelty:
Previous Methodology used gamma distribution as base distribution for
stochastic distances between region growing pixel boundaries.
In that, they defined the weight function as linear variation between
defined probabilities which restrict its adaptivity of different SAR intensity
images.
My contribution is that I changed the weight function to non-linear
according to the intensity of image which increase its adaptivity to a
certain extent.
7. Background
In Non-Local Means approach, based on stochastic distances for intensity speckle
reduction. A window is defined around each pixel, overlapping samples are compared
and only those which pass a goodness-of-fit test are used to compute the filtered
value. The noise-free estimated value of a pixel is defined as a weighted mean of
pixels in a certain region. Initially Stochastic Distances Nagao-Matsuyama approach is
considered in which neighbour hood is defined by Nagao-Matsuyama and samples are
compared and only those which pass a goodness-of-fit test based on stochastic
distances between distributions. Second is the referenced approach which is defined
later.
Assumptions:
Z = XY where Z is observed intensity, X is mean radar reflectivity, Y models speckle
noise
X follows Gamma Distribution
10. Comparison
Comparison is done using four indexes/measures
• SAR image indexes
o ENL(Equivalent number of looks)
o Edge saving index
• Optical Index
o Beta-Correlation
o Q-Correlation
Methods Probability ENL Edge Saving
Index
Beta-
Correlation
Q-
Correlation
Proposed 0.4 1.4892 0.4322 0.7106 2.057*e+03
Initial 0.4 1.4930 0.4314 0.7094 2.055*e+03
Proposed 0.5 1.7723 0.3758 0.6865 2.063*e+03
Initial 0.5 1.7802 0.3744 0.6847 2.060*e+03
Proposed 0.6 2.0849 0.3009 0.6429 1.983*e+03
Initial 0.6 2.0993 0.2983 0.6407 1.979*e+03
Given ENL for patch 50x50 is 0.5985, so proposed filter is more closer than previous filter
11. Conclusion
• This approach works for heterogeneous images
• Given no dependency to computational efficiency, this approach
works better than previous approaches
• All indexes estimate better results or in favour of new approach
• By doing stretching/increasing contrast of image in a certain
region, we can increase the indexes values
• Application : This can be used as a pre-processing step for
classification of different kind of heterogeneous images
Data Used: L-band UAV SAR
12. Future Work :
• Supervised Classification methods are used to classify the patches on
basis of their statistical parameters and hence, this will decrease the
computational time.
• Using Look-Up table approach, computational time can be reduced.