morphological algorithms for land mine detection · morphological algorithms for land mine...
Post on 19-Apr-2018
223 Views
Preview:
TRANSCRIPT
Morphological Algorithms for Land Mine Detection
Sinan Batman, Ulisses Braga-Neto and John Goutsias
Center for Imaging Science
The Johns Hopkins University
Baltimore, MD 21218
The second method, referred to here as MM-MNF (Fig. 4), is based on a
different linear prefilter: the Maximum Noise Fraction (MNF) transform.
The MNF transform is an optimal linear filter that maximizes the signal to
noise ratio (SNR) in each transform band subject to orthogonality. The
results from this linear stage is again submitted to the morphological
detection module shown in Fig. 5.
Two hybrid algorithms,
which combine the
decorrelating qualities of
a linear filter and the
shape extracting
properties of
Mathematical
Morphology, are
investigated in the
framework of land mine
detection.
The first algorithm, referred to here as PC-MM (Fig. 3), solely operates on
the peaks in the image that are extracted by a morphological top-hat
transform. These excerpted multi-spectral peaks are then compressed
and decorrelated via the principal component (PC) transform. Due to the
packing property of the PC transform, the target markers are typically found
in the first or second bands in the PC transformed image. The targets are
then detected using a morphological detection scheme (Fig. 5).
For the target signatures to be mapped to the
band with the highest (SNR), the MM-MNF
algorithm requires an accurate estimation of the
clutter covariance. This is achieved by using
the more stable PC-MM algorithm in a first pass.
The extracted targets from this first pass, are
then used to improve the detection result in
subsequent iterations using the MM-MNF
algorithm, by updating covariance estimates of
relevant filter variables (Fig. 6).
Unmanned AerialVehicle (UAV)
GPS Satellite
Field of View
Land Mines
Figure 1. Multi-spectral aerial imaging of land mines.
Band 1
Band 6Band 5Band 4
Band 3Band 2
Figure 2. Gray scale intensity profiles of multi-spectral image bands.
Top-hat byReconstruction
HistogramStretching Thresholding
Openingby
Reconstruction
Majority
Voting
Morphological
Reconstruction
Bandwise
Union
Masks
R
Structuring element Threshold level Structuring elementCut-off parameter
Morphological Detection (Direct MM Algorithm)
Figure 5.
Multispectral
Signal Ma
rke
rs
Binary
Detection
Result
The MM-MNF Algorithm
Opening byReconstruction
MNFTransform
Structuring element
Multispectral
Signal
BandSelection(Band 6)
Direct MMDetection
R
Binary
Detection
ResultClutter
Approximation
Figure 4.
PC-MM
MM-MNF
R1
MM-MNF
R3
R2
R2
R1
Detection Detection Image
After Binarizarion
Figure 6. Iterative implementation of Morphological Detection scheme
Multispectral
SignalMissed Target
Correct Detection
False Alarm
Iteration 1
Iteration 3
Iteration 2
Top-hat byReconstruction
PrincipalComponent
Decomposition
Structuring element
BandSelectionBand 1 orBand 2
Direct MMDetection
R
The PC-MM Algorithm
Multispectral
Signal
Binary
Detection
Result
Figure 3.
top related