automatic target recognition demonstration for cst reviewjao/talks/csttalks/atrdemo... ·...
TRANSCRIPT
Automatic Target Recognition Automatic Target Recognition Demonstration for CST ReviewDemonstration for CST Review
Professor Joseph A. OProfessor Joseph A. O’’SullivanSullivanLee Lee Montgnino Montgnino
Center for Security Technologies, Washington [email protected]://essrl.wustl.edu/~jao
Supported by: ONR, ARO, DARPAONR, ARO, DARPA
• Object Recognition and the Role of Templates• Our Methodology Based on Likelihoods• Comparative Results: MSTAR website• Open Problems:
– Fundamental Performance Bounds– Extensions to Optical Imagery
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
2
CollaboratorsCollaborators
Michael D. DeVoreNatalia A. SchmidLee MontagninoSushil AnandAndrew LiVikas Kedia
Donald L. SnyderDaniel R. FuhrmannMichael I. MillerJeffrey H. Shapiro
Faculty Students and Post-Docs
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
3
MotivationMotivation• Many reported approaches to ATR from SAR• Performance and database complexity are interrelated• We seek to provide a framework for comparison that:
- Allows direct comparison under identical conditions- Removes dependency on implementation details
2S12S1 T62T62 BTR 60BTR 60 D7D7 ZIL 131ZIL 131 ZSU 23/4ZSU 23/4
Publicly available SAR data from the MSTAR program
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
4
LEE THINKS: More MotivationLEE THINKS: More Motivation• Say we have MSTAR since it is readily available to all
•Experiments are performed in a controlled manner•Data is well understood, etc.
• Extensions into Optical imaging• Direct Link (?) to Airport Security
•The Scanners implemented in SEATAC• Other Modalities
Publicly available SAR data from the MSTAR program
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
5
Target Orientation EstimationTarget Orientation Estimation
Likelihood Model Approach:Likelihood Model Approach:ppRR||ΘΘ,,aa((rr||θθ,,aa) ) -- Conditional Data ModelConditional Data ModelppΘΘ,,aa((θθ,,aa) ) -- Prior on orientation (known or simply uniform)Prior on orientation (known or simply uniform)P(P(aa) ) -- Prior on target class (known or simply uniform)Prior on target class (known or simply uniform)
Target Target ClassifierClassifier ââ=T72=T72
Orientation Orientation EstimatorEstimator θθ=135=135
aa=T72=T72
Given a SAR image Given a SAR image rr, , determine a corresponding determine a corresponding target class target class ââ∈∈AA
Given a SAR image Given a SAR image rr and and a target class a target class ââ∈∈AA, , estimate target orientationestimate target orientation
Target Classification ProblemTarget Classification Problem
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
6
Training and Testing Problem:Training and Testing Problem:Function Estimation and ClassificationFunction Estimation and Classification
FunctionEstimation
L(r|a,θ) Inferenceââ=T72=T72
Scene and SensorPhysics
Training Data
Raw DataProcessing
Image
• Labeled training data: target type and pose• Log-likelihood parameterized by a function:
mean image, variance image, etc.
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
7
Function Estimation and ClassificationFunction Estimation and Classification
•• Functions are estimated fromFunctions are estimated from-- sample data sets onlysample data sets only-- physical model datasets only (PRISM, XPATCH, etc.)physical model datasets only (PRISM, XPATCH, etc.)-- combination of thesecombination of these
•• Training sets are finiteTraining sets are finite•• Computational and likelihood models have a finite Computational and likelihood models have a finite
number of parametersnumber of parameters•• Estimation errorEstimation error•• Approximation errorApproximation error•• Some regularization is neededSome regularization is needed
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
8
ApproachApproach• Select 240 combinations of implementation parameters• Execute algorithms at each parameterization• Scatter plot the performance-complexity pairs• Determine the best achievable performance at any complexity
BMP2 Variance Image at 6 Sizes
ZIL131 Variance Image at 6 Sizes
Six different image sizes from 128x128 to 48x48
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
9
System Parameters and ComplexitySystem Parameters and ComplexityApproximate Approximate αα((θθ,,aa) and ) and σσ22((θθ,,aa) as piecewise constant in ) as piecewise constant in θθImplementations parameterized by:Implementations parameterized by:
ww -- number of constant intervals in number of constant intervals in θθdd -- width of training intervals in width of training intervals in θθNN22 -- number of pixels in an imagenumber of pixels in an image
Database complexity ≡ log10(# floating point values / target type)
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
10
Performance and ComplexityPerformance and ComplexityForty combinations of angular resolution and training interval width.
Variance image of aT62 tank1 Window trained over 360°
Variance images of a T72 tank72 Windows trained over 10°
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
11
Problem StatementProblem Statement• Directly compare conditionally Gaussian, log-magnitude MSE,
and quarter power MSE ATR Algorithms- identical training and testing data- identical spatial and orientation windows
• Plot performance vs. complexity- probability of classification error- orientation estimation error- log-database size as complexity
• Use 10 class MSTAR SAR images
ApproachApproach• Select 240 combinations of implementation parameters• Execute algorithms at each parameterization• Scatter plot the performance-complexity pairs• Determine the best achievable performance at any complexity
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
12
Approaches: Conditionally Approaches: Conditionally GaussianGaussian
J. A. OJ. A. O’’Sullivan and S. Jacobs, IEEESullivan and S. Jacobs, IEEE--AES 2000AES 2000
Model each pixel as complex Model each pixel as complex GaussianGaussian plus uncorrelated noise:plus uncorrelated noise:
( ) ( )( )( )∏ +
−
Θ +=
i
NaKr
iA
i
i
eNaK
ap 0
2
,
0, ,
1, θ
θπθrR
ˆ a Bayes r( ) = argmaxa
maxk
ˆ p rθk ,a( )ˆ θ HS r,a( ) = argmax
θ k
ˆ p rθk ,a( )
GLRT Classification and MAP Estimation:GLRT Classification and MAP Estimation:
J. A. OJ. A. O’’Sullivan, M. D. Sullivan, M. D. DeVoreDeVore, V. , V. KediaKedia, and M. Miller, IEEE, and M. Miller, IEEE--AES to appear 2000AES to appear 2000
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
13
Approaches: LogApproaches: Log--MagnitudeMagnitudeMinimize distance between rdB = 20 log |r| and dB templates
d2 rdB,µLM( )= rdB − µLM2
Make decisions according to:
ˆ a LM r( )= argmina
mink
d2 rdB,µLM θk, a( )( )ˆ θ LM r a( )= argmin
θk
d2 rdB,µLM θk ,a( )( )
Alternatively, use a form of normalization:
d2 rdB − rdB,µLM θk,a( )− µLM θk ,a( )( )G. Owirka and L. Novak, SPIE 2230, 1994
L. Novak, et al., IEEE-AES, Jan. 1999
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
14
Approaches: Quarter PowerApproaches: Quarter Power
d2 rQP,µQP( )= rQP − µQP
2
Minimize distance between rQP = |r|1/2 and quarter power templates
Make decisions according to:
ˆ a QP r( ) = argmina
mink
d2 rQP,µQP θk, a( )( )ˆ θ QP r,a( ) = argmin
kd2 rQP,µQP θk,a( )( )
d2 rQP
rQP,
µQP θk ,a( )µQP θk ,a( )
⎛ ⎝ ⎜ ⎞
⎠
Or, normalized by vector magnitude:
S. W. Worrell, et al., SPIE 3070, 1997Discussions with M. Bryant of Wright Laboratory
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
15
PerformancePerformance--Complexity LegendComplexity Legend
Forty combinations of number of piecewise constant intervals and training window width
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
16
Conditionally Conditionally GaussianGaussian ResultsResults
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
17
Normalized Conditionally Normalized Conditionally Gaussian Gaussian ResultsResults
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
18
LogLog--Magnitude ResultsMagnitude Results
Recognition without normalization Arithmetic mean normalized
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
19
Quarter Power ResultsQuarter Power Results
Recognition without normalization Recognition with normalization
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
20
SideSide--byby--Side ResultsSide ResultsComparison in terms of:• Performance achievable at a given complexity• Complexity required to achieve a given performance
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
21
2S1 BMP 2 BRDM 2 BTR 60 BTR 70 D7 T62 T 72 ZIL131 ZSU 23 42S1 262 0 0 0 0 0 4 8 0 0 95.62%BMP 2 0 581 0 0 0 0 0 6 0 0 98.98%BRDM 2 5 3 227 1 0 14 3 5 4 1 86.31%BTR 60 1 0 0 193 0 0 0 0 0 1 98.97%BTR 70 4 5 0 0 184 0 0 3 0 0 93.88%D7 2 0 0 0 0 271 1 0 0 0 98.91%T 62 1 0 0 0 0 0 259 11 2 0 94.87%T 72 0 0 0 0 0 0 0 582 0 0 100%ZIL131 0 0 0 0 0 0 2 0 272 0 99.27%ZSU 23 4 0 0 0 0 0 2 0 1 0 271 98.91%
•• Probability of correctProbability of correctclassification: 97.2%classification: 97.2%
Target ClassificationTarget ClassificationResultsResults
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
22
• General problem in training/testing posed as estimation/classification
• Method of sieves (polynomial splines chosen)
• Comprehensive performance-complexity study for
- Ten class MSTAR problem
- Conditionally Gaussian model
- Log-magnitude MSE
- Quarter power MSE
• Provided a framework for direct comparison of alternatives and selection of implementation parameters
• Analysis ongoing
ConclusionsConclusions
J. A. O’Sullivan. CST Review, 01/13/2003ATR Demo
23
• Extensions into other domains….
•Optical, etc.
LEE THINKS: ExtensionsLEE THINKS: Extensions