AAccurate and Robust Automatic Target RecognitionAutomatic Target Recognition Method for SAR Imagery withMethod for SAR Imagery with

SOM-based ClassificationSOM based ClassificationShouhei Kidera and Tetsuo Kirimoto

University of Electro-Communications, Tokyo, JAPANInternational Symposium on Remote Sensing (ISRS) 2012 and y p g ( )

International Conference on Space, Aeronautics and Navigation Electronics (ICASANE) 2012, Incheon, Korea, 11th, Oct., 2012

IntroductionMicrowave radar : Applicable to adverse weather or darkness

SAR(Synthetic aperture radar) : High-resolution imaging methodSAR(Synthetic aperture radar) : High resolution imaging method

Target recognition with SAR imagery : A great deal of experience is required because SAR image is definitely different from optical image ⇒ ATR(Automatic Target Recognition) method is in demand

Optical image SAR image

SAR

Optical image SAR image

Conventional ATR TechniquesConventional ATR Techniques

Neural Network based approachTraditional ATR methods for SAR imagery

Neural Network based approach ・[1] C. M. Pilcher et al., IEEE Trans. Aerosp. Electron. Syst., 2011⇒ Classification employing range profile dataClassification employing range profile data

・[2] M. Martorella, et al., IEEE Trans. Aerosp. Electron. Syst., 2011⇒ Exploiting ISAR images with full polarimetric datap g g p

Other ATR approach ・[3] Q. Zhao, IEEE Trans. Aerosp. Electron. Syst., 2001⇒ SVM (Support Vector Machine) based classification

P bl i t diti l th dInaccurate classification in the case of

strong noisy situations or observation angle errors

Problem in traditional methods

strong noisy situations or observation angle errors⇒ More robust ATR method is proposed here !

System Model・Mono-static radar system・Targets with arbitrary shapes g y p・Transmitted signal : Frequency sweeping (complex value)・SAR image generation : Back projection algorithm ・Binarization method: Otsu’s discriminant analysis method

SAR image

System Model・Mono-static radar system・Targets with arbitrary shapes g y p・Transmitted signal : Frequency sweeping (complex value)・SAR image generation : Back projection algorithm

SAR imageSAR image Binarized image

・Binarization method: Otsu’s discriminant analysis method

Binarization

Conventional Method(Neural Network based Learning)

l k d l 3 l ( idd O )Neural network model : 3 layer (Input, Hidden, Output) [ ]

1,,1 Nxx L=x : Binarized SAR image sequence

)1(,)( == mxmu jj

[ ]1

Neuron model of each layer 3 layer neural network

)()(

1)(

)(,)(

1m

⎟⎞

⎜⎛

=

∑−

muNj

jj

)3,2(,

)1()(exp1!

,

=

⎟⎟⎠

⎞⎜⎜⎝

⎛−−+ ∑

=

m

mumwi

iji

)3,2(, m

)(mu j : Value of j th neuron

)(, mw ji : Weight from i th to j th neuron

Training method: Back propagation algorithm

Conventional Method(Neural Network based Classifier)

Classification Principle : Assessing difference among neuron’s values of output layer

Optimal class number is determined :

)3(

)3()3(minarg

tr

tr

1

NNopt

k

NkK

u

uu −=

≤≤ )3(1 tr kNk u≤≤

: Output of training data)3(trku

)3(u : Output of test data

Problem : Inaccuracy in lower SNR situations or other observation errors (e.g. Angle of nose)

Proposed ATR Method(SOM Training phase)

SOM (S lf O i i M ) U i d l t k th d

Proposed scheme : SOM classification employing training data

SOM (Self Organizing Map): Unsupervised neural network method

⇒ Supervised SOM

Other feature Periodical structure of SOM

1. Torus SOM : Periodical map structure

Other feature

: Periodical map structure ⇒ Avoiding undesired bias of node

2 BLSOM (Batch Leaning SOM)2. BLSOM (Batch Leaning SOM): Impervious to the order of training sequence

Proposed ATR Method(Actual SOM training procedure)I iti l t t t d i d fi d( ) ∑∑ ==

= trtr

11tr )()(0; N

k kN

k kk aa pxppyInitial output vector on node p is defined p : Location of node

trkx : Training SAR image

( ) tr;maxarg)(ˆ tt xpyp =For k th tranning data , winner node is determined)(ˆ tkptr

kx( );maxarg)( kk tt xpyp

p−=

Ω∈

( ) ( ) ( )( )∑ −tr

1tr ;)),(ˆ(N

k kk tth pyxpp

After calculating for all training data, The output of each node is updated by :

)(ˆ tkp

( ) ( ) ( )( )∑

∑=

=+=+tr

1

1

)),(ˆ(

;)),((;1; N

k k

k kk

thtt

pp

pypppypy

⎟⎞

⎜⎛ −)(ˆ

ˆtk pp )()( tt σβ : Monotonically

⎟⎟⎠

⎞⎜⎜⎝

⎛= 2)(2

)(exp)()),(ˆ(

tt

tth kk σ

βpp

pp )(),( tt σβ : Monotonically decreasing for t

Proposed ATR Method(Classification Phase)

Cl ifi i P i i lClassification Principle : Assessing value of integral of U-matrix field from training node

( ) xpyxp −= som;minarg)(ˆ TWinner node for test data x

U-matrix field on SOM

p Ω∈

Optimal class number is determined :

{ }∫≤≤

=),(),(1

somopt d)(minminarg

trxx

pkCkCNk

sUK

)( pU : U-matrix potential at node p),( xkC : Possible path from k th training node ),( p g

to winner node )(ˆ xp

Proposed ATR Method(Classification Phase: U-matrix metric)

Cl ifi i P i i lClassification Principle : Assessing value of integral of U-matrix field from training node

( ) xpyxp −= som;minarg)(ˆ TWinner node for test data x

p Ω∈

Optimal class number is determined :

{ }∫≤≤

=),(),(1

somopt d)(minminarg

trxx

pkCkCNk

sUK

)( pU : U-matrix potential at node p),( xkC : Possible path from k th training node ),( p g

to winner node )(ˆ xp U-matrix field

Procedure of Proposed MethodProcedure of Proposed Method

BLSOM

Experimental Validation Experimental setup :1/200 downscaled model of X-band radar

except for center frequency

・Horn antennas (Beamwidth : 27 deg) ・Frequency range: 24GHz – 40GHz ・Slant range resolution : 9.375 mm ・Aperture Length : 1600 mmSlant range resolution : 9.375 mm Aperture Length : 1600 mm・Off-nadir angle : 54.7 degree ・Tx and Rx Separation : 48 mm

Optical image for 5 civilian airplanes Observation scene in experimentB747 B787

Optical image for 5 civilian airplanes Observation scene in experiment

A320DC-10B777

Evaluation in Noisy Situationva ua o No sy S ua oGaussian noises are numerically

Correct classification probabilityy

added to SAR images for contaminated image generation

U-matrix potential distribution

Robustness to Observation Angle ErrorsRobustness to Observation Angle ErrorsAngle observation error : φ

φφ = 0 deg

Robustness to Observation Angle ErrorsRobustness to Observation Angle ErrorsAngle observation error : φ

φφ = 15 deg

ConclusionConclusion• Accurate ATR method based on Supervised SOM• Accurate ATR method based on Supervised SOM

has been proposed P d th d・Proposed method1. Supervised SOM for ATR classification issue2 New classification metric by using U-matrix metric2. New classification metric by using U-matrix metric

・Experimental validation : ・Correct classification even in under SNR=10 dB・Robust feature for observation angle errors

Future work ・More accurate method exploiting complex value of SAR imageMore accurate method exploiting complex value of SAR image