air - northeastern universityt5: roaming ex-plosives detection sensor suite luggage fieldable...
TRANSCRIPT
Robust Multifrequency Inversion in Terahertz TomographyKe Chen and David Castanon ECE Department, Boston University
Motivation• Problem: The developed joint multifrequency inversion technique out-
performs single-frequency inversion, but requires complete and accuratespectral information
• Goal: Robust reconstruction approaches which do not assume exact priorinformation of spectra
Research to realityThis work provides a general framework for robust joint image formationusing multiple frequencies for improved reconstruction and sequent recogni-tion. The algorithms are applicable to Terahertz(THz) pulsed imaging whichhas attracted much attention for its ability to reveal unique spectral charac-teristics of chemicals in THz range and thus to fingerprint explosives.
Background
Object function:o(r, f)= n(r, f)2−1
Measurement:us(r)=u(r)−u0(r)
• Fourier Diffraction Theorem[1] relates scattering with object spectrum un-der the Born approximation u(r)≈u0(r)
Us(ωx) =ieilωy
2ωyO(ωx, ωy − k0), ωy ,
√k20 − ω2
x
• Model: {fm} in THz radiation spectrum
ym = Ψmxm, m = 1, . . . ,M
ym ∈ CKm Fourier transform(FT) of THz field measurement
xm ∈ CN Estimate of object function at fm
Ψm Nonuniform FT operator with fast approximation NUFFT[2] Tm ≈ Ψm
Joint inversion with E
JM: min(x,s)‖y − T x‖
2
+ α21‖DMx‖
2
WM+ ϕ(s, γ)
where x = [x1 , . . . ,xM]T , T , diag[Tm], Wn := W⊗ In, derivative operatorD
• Data-fidelity term in frequency domain, smooth penalty in spatial domain• Jointly construct boundaries s ∈ RN : Invariant across {fm}• Mumford-Shah[3]: ϕ(s, γ)=γ2‖∇s‖2 + 1
γ2 ‖s‖2
Spatially varying weighting: W , diag[(1− [s]i)
2]
Methodologies
Joint inversion with EJMP: min
(ν,s)‖y − T H ν‖
2
+ α21‖DJν‖
2
WJ+ ϕ(s, γ)
where ν=[ν1,...,ν
J]T ,H=(E1···EJ )⊗IN
• Given exact spectral prior Ej ∈ CM , j = 1, . . . , J
• Abundance of the j-th material νj ∈ RN , [νj ]i∈ [0, 1]
• Linear transform[4] x = Hν
Case 1. With partial E
RJMP1: min(x,s)‖y − T x‖
2
+ α21‖D(M+J)x‖
2
W(M+J)+ α2
2‖xc‖1
1+ ϕ(s, γ)
where xc=[xc1, . . . ,x
cM ]T, x=[ν,xc]T , T =(T H, T )
• Augment xcm∈ CN to compensate prior incomplete
• `1norm penalty on xc: ‖x‖11≈∑Ni=1
([x]2i + β
)1/2Case 2. With uncertain E• Uncertainty modelling: Ej−∆j ≤ ej ≤ Ej+∆j
• Inverse with worst-case prior: H = (e1 · · · eJ)⊗ IN
RJMP2: min(ν,s)
ψ(ν, s) + α21‖DJν‖
2
WJ+ ϕ(s, γ)
ψ(ν, s) , maxH‖y − T H ν‖
2
• NP-hard solving ψ: possible solution ψj(ν, s) at Hj , j=1, . . . , 22J
• Smoothing approximation[5]: ψp(ν, s) , λ−1log(∑
j exp(λψj(ν, s))
Experiment
Target and spectrum prior
An 81×81 phantom consisting of 3 explo-sives and air as background was generatedwith known spectral prior[6]. We simulatedTHz incident fields at 19 projection angles,and collected complex amplitude of the scat-tering at 4 frequencies for reconstruction.Each method was evaluated under variouslevels of Gaussian noise.
Noiseless SNR30 SNR20 SNR10
0.05
0
0.01
0.02
0.03
0.04 JM
JMP
1.8THz
wc-JMP
RJMP2
RJMP1
Noiseless SNR30 SNR20 SNR10
0.05
0
0.01
0.02
0.03
0.04JM
JMP
2.0THz
wc-JMP
RJMP2RJMP1
Noiseless SNR30 SNR20 SNR10
0.04
0
0.01
0.02
0.03JM
JMP
2.4THz
wc-JMP
RJMP2RJMP1
Noiseless SNR30 SNR20 SNR10
0.04
0
0.01
0.02
0.03 JM
JMP
3.0THz
wc-JMP
RJMP2
RJMP1
Exp1. ReconstructionRJMP1 is given only spec-tra of RDX. wc-JMP andRJMP2 are given the worst-case spectral priors under5% uncertainty.
Noiseless SNR30 SNR20 SNR10 SNR5
0.09
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08 JM
RJMP1
JMP
RDX
Noiseless SNR30 SNR20 SNR10 SNR5
0.09
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
JM
RJMP1JMP
PETN
Noiseless SNR30 SNR20 SNR10 SNR5
0.09
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
JM
RJMP1
JMP
Metabel
Noiseless SNR30 SNR20 SNR10 SNR5
0.01
0
0.001
0.002
0.004
0.006
0.008
JM
Air
Air
JMP
Exp2. Recognitionpixel-wise labeling j at r:JM:minimizer of Euclideandistance from x to Ej .JMP: maximizer νj .RJMP1: maximizer νj forsmall xc
Discussion
• JMP: force spectral and spatial consistency during inversion; sensitive tothe completeness and accuracy of spectral priors
• RJMP1: encourage sparse representation using given spectral priors
• RJMP2: : minimax optimization techniques
• Future work: handle uncertainties in accuracy and completeness at thesame time
References
ALERT
Three-Level Strategic Approach
Level 3: Grand
Challenges
Level 2
Enabling
Technology:
Testbeds
Level 1
Fundamental
Science
C1: Ultra-Reliable
Screening
C3: Unequivocal Pre-
& Post-Blast
Mitigation
C2: >50 meter
Stand-Off Discovery
and Assessment
C4: Rapid & Thorough
Preparedness and
Response
F1: Explosives
Characterization
F2: Explosives
Detection Sensors
F3: Explosives
Detection Sensor
Systems
F4: Blast
Mitigation
T1: Multimodal
Suicide Bomber
Detection
T2: Trace
Explosives
Detection/Screening
T3: VBIED
Signatures
Testbed T5: Roaming Ex-
plosives Detection
Sensor Suite
T4: Multisensor
Luggage
Scanner
Fieldable
Products
Teaming with
National Labs
& Industry
Research
Drivers
Research
Barriers
Proof-of-Principle
Tests of Research
This Work
[1] A. Kak, M. Slaney, Principles of Computerized Tomographic Imaging, 2001.[2] J. Fessler, B. Sutton, IEEE Tr. Sig. Proc. vol. 51, no. 2, pp. 560-574, 2003.[3] R. Weisenseel, Ph.D. diss., Boston University, 2004.[4] A. Li and et al. Appl. Opt. vol. 44, no. 10, pp.1948-1956, 2005.[5] E. Polak and et al. J Optimiz Theory App vol. 119, no. 3, 2003.[6] C. Baker and et al. Proc. of IEEE vol.95, no. 8, pp1559-1565, 2007.
This work was supported in part by Gordon-CenSSIS, the Bernard M. Gordon Center for Subsurface Sensing andImaging Systems, under the Engineering Research Centers Program of the National Science Foundation (Award NumberEEC-9986821), and the U.S. Department of Homeland Security (award number 2008-ST-061-ED0001).
1