Beruflich Dokumente
Kultur Dokumente
Measurement Uncertainty
Seung Ho Doo1, Graeme Smith,1 and Chris Baker1
1
Electrical and Computer Engineering
The Ohio State University
Columbus, OH, USA
{doo.3, smith.8347, baker.1891}@osu.edu
I. INTRODUCTION
In this paper, target classification performance of the
MSTAR dataset is examined as a function of measurement
uncertainty using specially designed features that both cope
with and exploit signature variability. Performance is reported
as a function of measurement uncertainty conditions [1].
The MSTAR dataset is one of the most widely used real
target SAR images in Automatic Target Recognition (ATR)
community [2-4]. Various kinds of target classification
techniques have been developed for a MSTAR target ATR
system, and significantly high classification performance,
which is over 90%, is reached by some publications [2,5]. This
high classification performance could be acquired not only by
the high performance of the target classification systems, but
also the idealness of the MSTAR dataset. In other words, the
Signal to Noise Ratio (SNR) of the MSTAR dataset is high,
over 30 dB, and the target alignment also is nearly perfect.
However, in real target classification situations, more difficult
conditions can occur such as low SNR, target aspect angle
estimation failure, and inconsistent elevation angles between
test and training targets.
In [1], we proposed reliable target feature extraction
techniques which can cope with the variability contained in
target measurements. The proposed features include dominant
scattering area size, dominant scattering area, and scatterer
positional stability information. The proposed features are
c
978-1-4673-7297-8/15/$31.00 2015
IEEE
587
0.591 GHz
Center Frequency
9.59 GHz
Full 360
Down-Range Resolution
0.3047 m
Cross-Range Resolution
0.3047 m
Target Range
Approximately 5 Km
Elevation Angles
Angular Spacing
Approximately 0.03
Data type
SAR image
588
(1)
TABLE II. TARGET CLASSIFICATION RESULTS BY THREE PROPOSED FEATURES (DSA, DSAS, AND SPSI)
T72
BTR60
BMP2
2S1
ZSU23
BRDM2
BTR60
D7
T72
ZIL131
Average
Rate
T72
0.980
0.000
0.020
0.011
0.007
0.018
0.005
0.000
0.007
0.000
BTR60
0.000
0.980
0.005
0.004
0.000
0.011
0.021
0.000
0.007
0.000
BMP2
0.020
0.000
0.974
0.007
0.004
0.011
0.010
0.004
0.004
0.000
2S1
0.000
0.000
0.000
0.945
0.004
0.000
0.000
0.004
0.007
0.007
ZSU23
0.000
0.000
0.000
0.011
0.971
0.000
0.000
0.004
0.015
0.004
BRDM2
0.000
0.015
0.000
0.004
0.004
0.905
0.005
0.000
0.004
0.000
BTR60
0.000
0.005
0.000
0.000
0.000
0.026
0.949
0.000
0.004
0.007
D7
0.000
0.000
0.000
0.000
0.007
0.004
0.000
0.985
0.004
0.000
T72
0.000
0.000
0.000
0.000
0.004
0.011
0.000
0.004
0.901
0.004
ZIL131
0.000
0.000
0.000
0.018
0.000
0.015
0.010
0.000
0.048
0.978
0.957
Fig. 1. Target classification result with varying aspect angle estimation error
Fig. 2. For example, in the case of a tank target, the direction of a turret can
vary, so the test and training targets have different
both test and training data. As the SNR changes, the mean of
the correct target classification ratio is examined. The SNR of
target measurements is calculated as follows.
(2)
where is a variance operator. Since it is not possible to
perfectly delete the noise from the MSTAR SAR images, it is
assumed the variance of image in equation (2) is calculated
directly from the MSTAR SAR images.
The SNR changes from 50 dB to -10 dB, and the 10 target
classification results are examined. Fig. 2 shows the target
classification ratio with varying SNR. As can be seen, the
classification ratio with noise contaminated SAR image is over
0.9 under the strong noise, -4 dB SNR. This shows the
robustness of the proposed features under strong noise
classification environments.
D. Different Test and Training Data
In real target classification, it is not practical to assume that
the test and training targets are exactly identical to each other.
BMP2 (Test/Training)
9566 / C21
9566 / C21
9566 / C21
9563 / 9566
9563 / 9566
9563 / 9566
9566 / C21
9566 / C21
9566 / C21
T72 (Test/Training)
132 / S7
132 / 812
812/ S7
132 / S7
132 / 812
812/ S7
132 / S7
132 / 812
812/ S7
589
TABLE III. TARGET CLASSIFICATION RESULTS OF DIFFERENT TEST AND TRAINING TARGETS
T72 (S7)
BTR60
BMP2
(C21)
2S1
ZSU23
BRDM2
BTR60
D7
T72
ZIL131
Average
Rate
T72
(812)
0.754
0.000
0.016
0.985
BMP2
(9563)
0.136
0.000
0.056
0.020
0.842
0.056
0.010
0.010
0.005
0.000
0.000
0.000
0.007
0.004
0.018
0.005
0.004
0.011
0.000
0.007
0.000
0.011
0.021
0.000
0.004
0.000
0.007
0.004
0.007
0.000
0.004
0.004
0.000
0.938
0.000
0.000
0.000
0.000
0.007
0.007
0.011
0.978
0.000
0.000
0.004
0.015
0.000
0.004
0.004
0.901
0.005
0.000
0.004
0.000
0.007
0.000
0.026
0.964
0.000
0.004
0.007
0.000
0.007
0.004
0.000
0.985
0.004
0.000
0.000
0.004
0.015
0.000
0.004
0.901
0.007
0.018
0.000
0.018
0.005
0.000
0.048
0.978
BTR60
2S1
ZSU23
BRDM2
BTR60
D7
T72
ZIL131
0.010
0.000
0.021
0.000
0.026
0.010
0.010
0.005
0.000
0.000
0.026
0.000
0.000
0.000
0.923
590