model-based image reconstruction of chemiluminescence...
TRANSCRIPT
Model-based image reconstruction ofchemiluminescence using a plenoptic 2.0 camera
Hung Nien, Jeffrey A. Fessler, and Volker Sick
EECS Department and Mechanical Engineering DepartmentUniversity of Michigan, Ann Arbor
ICIP 2015: Computational ImagingSept. 29, 2015
1 / 29
Disclosure
I Supported by NSF under grant number CBET-1402707
I Equipment support from Intel Corporation
2 / 29
Motivation: combustion in transparent engine cylinder
3 / 29
Motivation
Tomographic reconstruction of 3D chemiluminescence patternssuch as flame fronts using a plenoptic camera.
Previous workI Tomo-PIV (particle image velocimetry) (4–6 cameras) [Elsinga et al.,
2006]
I Plenoptic 1.0 camera for PIV [Fahringer et al., 2012]
I Single-camera stereo [Greene et al., 2013] [Chen et al., 2015]
Depth maps for translucent objects?
4 / 29
Plenoptic camera
Plenoptic cameras use micro-lens arrays to capture 4-D light fieldinformation of a scene. The angular information enables:
I depth estimation (for object surfaces illuminated externally)e.g., via triangulation [Perwaß, SPIE, 2012]
I tomographic reconstruction (for luminescent objects)(cf., digital X-ray tomosynthesis - limited-angle tomography).
� Images courtesy of Raytrix GmbH and Lytro, Inc.
5 / 29
Model-based image reconstruction (MBIR)
Overall goal: reconstruct 3D chemiluminescence pattern x fromplenoptic camera measurement y.
MBIR components:I 3D object model (basis coefficients) x
- Image voxel, basis function, ...
I System model A (# of sensor elements × # of object voxels)
- Linearity, finite voxel size, finite pixel size, ...
I Data noise statistics p(y|Ax)
- Additive Gaussian, Poisson, ...
I Cost function Ψ(x)
- Data fidelity, regularizer, physical constraints, ...
I Iterative algorithm (arg minx)
- MART, FISTA, Newton’s methods, ...
[Nuyts et al., Phys. Med. Biol., 2013]
6 / 29
Model-based image reconstruction (continued)
We reconstructed objects by solving a regularized LS problem:
x? ∈ arg minx
{Ψ(x) , 1
2 ‖y − Ax‖22 + R(x)
}s.t. x � 0 ,
where R denotes an edge-preserving corner-rounded TV regularizer.
We focused on R defined as
R(x) ,∑
i=1 βi∑
n ϕHuber([Cix]n) ,
I Ci : finite difference matrix along ith direction
I βi : corresponding regularization parameter.
I ϕHuber(t) ≈ |t|
7 / 29
System model for a plenoptic camera
[Bishop & Favaro, IEEE T-PAMI, 2012]
8 / 29
System model for a plenoptic camera
Build (pre-compute) system matrix A one column at a time
[Bishop
& Favaro, IEEE T-PAMI, 2012]
8 / 29
System model for a plenoptic camera
| ←− z −→ | ←− Z −→ |thin lens formula: 1/z + 1/Z = 1/F
[Bishop & Favaro, IEEE T-PAMI,
2012]
8 / 29
System model for a plenoptic camera
[Bishop & Favaro, IEEE T-PAMI, 2012]
8 / 29
System model for a plenoptic camera
[Bishop & Favaro, IEEE T-PAMI, 2012]
8 / 29
System model for a plenoptic camera
Use superposition to consider one microlens at a time
[Bishop &
Favaro, IEEE T-PAMI, 2012]
8 / 29
System model for a plenoptic camera
If microlens had large diameter...
[Bishop & Favaro, IEEE T-PAMI,
2012]
8 / 29
System model for a plenoptic camera
If main lens had large diameter...
[Bishop & Favaro, IEEE T-PAMI,
2012]
8 / 29
System model for a plenoptic camera
Combined effect of main lens and microlens
[Bishop & Favaro, IEEE
T-PAMI, 2012]
8 / 29
System model for a plenoptic camera
Combined effect of main lens and microlens
[Bishop & Favaro, IEEE
T-PAMI, 2012]
8 / 29
System model - continued
Continuous-space PSF of the ith micro-lens is:
βi (s, t; x , y , z) = βML-µLi (s, t; x , y , z)︸ ︷︷ ︸∝ circ
(s,t;cML-µL
i ,Bi
) ·βµLi (s, t; x , y , z)︸ ︷︷ ︸∝ circ
(s,t;cµL
i ,bi
) ,where
I (s, t) denotes 2D sensorcoordinates
I centers cML-µLi , cµL
i , and radii Bi ,
and bi depend on the object
point position (x , y , z) and
camera geometry.
I∑
i βi (s, t; x , y , z) sketched:
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
s−axis [mm]
t−axis
[m
m]
Continuous−space PSF
9 / 29
Computational challenges
I Dense micro-lens array
I Highly shift-variant point spread function
I Non-separable aperture / PSF (cf., X-ray CT)
I Lens aberrations
I Finite sensor pixel sizeThe discrete PSF of a micro-lens consists of integrals of the circle-circleintersection over each sensor pixel, where the circle centers depend on theposition of the “point source.”
We approximate each finite-sized sensor pixel as L× L infinitesimal pixels,
i.e., L×-subsampling in each direction.
I Finite object voxel size
10 / 29
Computational challenges
I Dense micro-lens array
I Highly shift-variant point spread function
I Non-separable aperture / PSF (cf., X-ray CT)
I Lens aberrations
I Finite sensor pixel sizeThe discrete PSF of a micro-lens consists of integrals of the circle-circleintersection over each sensor pixel, where the circle centers depend on theposition of the “point source.”
We approximate each finite-sized sensor pixel as L× L infinitesimal pixels,
i.e., L×-subsampling in each direction.
I Finite object voxel size
10 / 29
Finite-sized object voxel effects
One (x , y) transaxial plane of a 3D object voxel
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
9.8x 10
−5 PSF zoom
s [mm]t
[mm
]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
9.8x 10
−5
11 / 29
Finite-sized object voxel effects
One (x , y) transaxial plane of a 3D object voxel
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
9.8x 10
−5 PSF zoom
s [mm]t
[mm
]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
9.8x 10
−5
11 / 29
Finite-sized object voxel effects
One (x , y) transaxial plane of a 3D object voxel
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
9.8x 10
−5 PSF zoom
s [mm]t
[mm
]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
9.8x 10
−5
11 / 29
Finite-sized object voxel effects
One (x , y) transaxial plane of a 3D object voxel
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
9.8x 10
−5 PSF zoom
s [mm]t
[mm
]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
9.8x 10
−5
11 / 29
Finite-sized object voxel effects
One (x , y) transaxial plane of a 3D object voxel
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
9.8x 10
−5 PSF zoom
s [mm]t
[mm
]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
9.8x 10
−5
11 / 29
Finite-sized object voxel: baseline approximation
We approximate each cubic voxel as K × K × K equally spacedinfinitesimal voxels, i.e., K×-subsampling in each direction.
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
4.8x 10
−3 PSF zoom
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
4.8x 10
−3
† We used K = 7 here.
12 / 29
Finite-sized object voxel: baseline approximation
We approximate each cubic voxel as K × K × K equally spacedinfinitesimal voxels, i.e., K×-subsampling in each direction.
PSF
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
9.8x 10
−5 PSF zoom
s [mm]t
[mm
]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
9.8x 10
−5
(infinitesimal)
12 / 29
Why finite voxel size matters
(zoomed)
Infinitesimal voxels
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
0.285
Finite−sized voxels
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
0.285
† Using 0.5× 0.5× 0.5 [mm3] cubic voxels and K = 7.
13 / 29
Why finite voxel size matters (zoomed)
Infinitesimal voxels
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
0.285
Finite−sized voxels
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
0.285
Abs. difference [0 0.034]
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
0.035
† Using 0.5× 0.5× 0.5 [mm3] cubic voxels and K = 7.
13 / 29
Numerical experiments: imaging geometry
14 / 29
Numerical experiments: object geometry
I 100× 100× 100 voxel object
I 0.5× 0.5× 0.5 [mm3] voxels
I 50 [mm] field-of-view
I 7× sensor subsampling whenprecomputing A
I 50 dB SNR (additive whiteGaussian noise)
I To avoid an inverse crime when synthesizing plenoptic sensor pictures, weused a voxelized object having a 2× finer grid in 3D, with 11×subsampling per dimension.
15 / 29
Numerical experiments: object geometry
I 100× 100× 100 voxel object
I 0.5× 0.5× 0.5 [mm3] voxels
I 50 [mm] field-of-view
I 7× sensor subsampling whenprecomputing A
I 50 dB SNR (additive whiteGaussian noise)
I To avoid an inverse crime when synthesizing plenoptic sensor pictures, weused a voxelized object having a 2× finer grid in 3D, with 11×subsampling per dimension.
15 / 29
Numerical experiments: object geometry
I 100× 100× 100 voxel object
I 0.5× 0.5× 0.5 [mm3] voxels
I 50 [mm] field-of-view
I 7× sensor subsampling whenprecomputing A
I 50 dB SNR (additive whiteGaussian noise)
I To avoid an inverse crime when synthesizing plenoptic sensor pictures, weused a voxelized object having a 2× finer grid in 3D, with 11×subsampling per dimension.
15 / 29
Numerical experiments: plenoptic cameras
Camera #1 Camera #2 Camera #3
fmain 80 80 80 [mm]f-number 1.4 2.8 1.4dmain 57.14 28.57 57.14 [mm]
fmicro 0.35 0.35 0.35 [mm]dmicro 0.27 0.135 0.135 [mm]
type larger Bishop & Favaro overlaps
I 9µm × 9µm sensor pixel size
I 850× 850 pixel sensor
16 / 29
Numerical experiments: plenoptic cameras
Camera #1 Camera #2 Camera #3
fmain 80 80 80 [mm]f-number 1.4 2.8 1.4dmain 57.14 28.57 57.14 [mm]
fmicro 0.35 0.35 0.35 [mm]dmicro 0.27 0.135 0.135 [mm]
type larger Bishop & Favaro overlaps
I 9µm × 9µm sensor pixel size
I 850× 850 pixel sensor
16 / 29
Numerical experiments: plenoptic cameras
Camera #1 Camera #2 Camera #3
fmain 80 80 80 [mm]f-number 1.4 2.8 1.4dmain 57.14 28.57 57.14 [mm]
fmicro 0.35 0.35 0.35 [mm]dmicro 0.27 0.135 0.135 [mm]
type larger Bishop & Favaro overlaps
I 9µm × 9µm sensor pixel size
I 850× 850 pixel sensor
16 / 29
Numerical experiments: simulated plenoptic pictures
Camera #1
s [mm]
-3 -2 -1 0 1 2 3
t [m
m]
-3
-2
-1
0
1
2
3
0
2.48
Better angular resolution than Camera #2, but worse spatial resolution
17 / 29
Numerical experiments: simulated plenoptic pictures
Camera #2
s [mm]
-3 -2 -1 0 1 2 3
t [m
m]
-3
-2
-1
0
1
2
3
0
2.47
Bishop & Favaro, 2012
17 / 29
Numerical experiments: simulated plenoptic pictures
Camera #3
s [mm]
-3 -2 -1 0 1 2 3
t [m
m]
-3
-2
-1
0
1
2
3
0
1.37
Overlapping (larger) subimages: demultiplexing needed,
but (perhaps) more angular information than Camera #2.
17 / 29
Numerical experiments: image reconstruction
[Recap] Our image reconstruction problem is:
x? ∈ arg minx
{Ψ(x) , 1
2 ‖y − Ax‖22 + R(x)
}s.t. x � 0 ,
I 500 iterations of FISTA with adaptive restart[Beck & Teboulle, IEEE T-IP, 2009]
[O’Donoghue & Candes, FCM, 2015]
I Precomputed / stored A (column-wise sparse)
- 7× object subsampling (x , y , z)- 7× sensor subsampling (s, t)- 320 secs/slice for camera #1
(60 threads @ MATLAB)
I 26 3D neighbors used in the (smoothed) TV regularizer
18 / 29
Numerical experiments: finite voxel size
Infinitesimal voxels Finite-sized voxels
Contours at isovalue = 20% of maximum intensity.
19 / 29
Numerical experiments: aperture size
Large aperture/coarse MLA Small aperture/dense MLA
Better lateral resolution of Camera 2 not helpful here.
20 / 29
Numerical experiments: aperture size - slices
Phantom
xy yz xz
0
10
Camera #1 [approx., SNR = 50 dB]
xy yz xz
0
3Camera #2 [approx., SNR = 50 dB]
xy yz xz
0
3
21 / 29
Numerical experiments: overlapping subimages
Non-overlapping subimages Overlapping subimages
22 / 29
Numerical experiments: overlapping subimages - slices
Phantom
xy yz xz
0
10
Camera #2 [approx., SNR = 50 dB]
xy yz xz
0
3Camera #3 [approx., SNR = 50 dB]
xy yz xz
0
3
23 / 29
Numerical experiments: object distance
dobject = 700 [mm] dobject = 550 [mm]
24 / 29
Numerical experiments: object distance - slices
Phantom
xy yz xz
0
10
Camera #2 [approx., SNR = 50 dB]
xy yz xz
0
3Camera #2 [approx., SNR = 50 dB]
xy yz xz
0
3
dobject = 700 [mm] dobject = 550 [mm]
25 / 29
Numerical experiments: sharp vs smooth object
Sharp object edges Smooth object edges
26 / 29
Numerical experiments: sharp vs smooth object - slices
Sharp phantom
xy yz xz
0
10Smooth phantom
xy yz xz
0
10
Camera #3 [approx., SNR = 50 dB]
xy yz xz
0
3Camera #3 [approx., SNR = 50 dB]
xy yz xz
0
3
27 / 29
Conclusions
I Model-based image reconstruction may be viable for 3Dchemiluminescence from plenoptic camera data
I Voxel-size modeling is important
I Larger angular range of incident light improves z-resolution(but more severe lens aberration?)
I F-number matching can be relaxed (overlapping sub-images)to improve depth resolution in tomographic formulation
I Need fast on-the-fly forward/back-projections to solve reallarge-scale image reconstruction problems
28 / 29
Conclusions
I Model-based image reconstruction may be viable for 3Dchemiluminescence from plenoptic camera data
I Voxel-size modeling is important
I Larger angular range of incident light improves z-resolution(but more severe lens aberration?)
I F-number matching can be relaxed (overlapping sub-images)to improve depth resolution in tomographic formulation
I Need fast on-the-fly forward/back-projections to solve reallarge-scale image reconstruction problems
28 / 29
Conclusions
I Model-based image reconstruction may be viable for 3Dchemiluminescence from plenoptic camera data
I Voxel-size modeling is important
I Larger angular range of incident light improves z-resolution(but more severe lens aberration?)
I F-number matching can be relaxed (overlapping sub-images)to improve depth resolution in tomographic formulation
I Need fast on-the-fly forward/back-projections to solve reallarge-scale image reconstruction problems
28 / 29
Conclusions
I Model-based image reconstruction may be viable for 3Dchemiluminescence from plenoptic camera data
I Voxel-size modeling is important
I Larger angular range of incident light improves z-resolution(but more severe lens aberration?)
I F-number matching can be relaxed (overlapping sub-images)to improve depth resolution in tomographic formulation
I Need fast on-the-fly forward/back-projections to solve reallarge-scale image reconstruction problems
28 / 29
Conclusions
I Model-based image reconstruction may be viable for 3Dchemiluminescence from plenoptic camera data
I Voxel-size modeling is important
I Larger angular range of incident light improves z-resolution(but more severe lens aberration?)
I F-number matching can be relaxed (overlapping sub-images)to improve depth resolution in tomographic formulation
I Need fast on-the-fly forward/back-projections to solve reallarge-scale image reconstruction problems
28 / 29
Bibliography
[1] G. E. Elsinga, B. Wieneke, F. Scarano, and A. Schroder, “Tomographic 3D-PIV and Applications,” inParticle Image Velocimetry, Springer Berlin Heidelberg, 2008, pp. 103–25.
[2] T. W. Fahringer and B. S. Thurow, “Tomographic reconstruction of a 3-D flow field using a plenopticcamera,” in AIAA Fluid Dynamics Conf., 2012, p. 2826.
[3] M. L. Greene and V. Sick, “Volume-resolved flame chemiluminescence and laser-induced fluorescenceimaging,” Appl Phys B, vol. 113, no. 1, 87–92, Oct. 2013.
[4] H. Chen, P. Lillo, and V. Sick, “3D spray-flow interaction in a spark-ignition direct-injection engine,” Intl.J. Engine Research, 2015, To appear.
[5] C. Perwass and L. Wietzke, “Single lens 3D-camera with extended depth-of-field,” in Proc. SPIE 8291Human Vision and Electronic Imaging XVII, 2012, p. 829 108.
[6] J. Nuyts, B. De Man, J. A. Fessler, W. Zbijewski, and F. J. Beekman, “Modelling the physics in iterativereconstruction for transmission computed tomography,” Phys. Med. Biol., vol. 58, no. 12, R63–96, Jun.2013.
[7] T. E. Bishop and P. Favaro, “The light field camera: Extended depth of field, aliasing, andsuperresolution,” IEEE Trans. Patt. Anal. Mach. Int., vol. 34, no. 5, 972–86, May 2012.
[8] A. Beck and M. Teboulle, “Fast gradient-based algorithms for constrained total variation image denoisingand deblurring problems,” IEEE Trans. Im. Proc., vol. 18, no. 11, 2419–34, Nov. 2009.
[9] B. O’Donoghue and E. Candes, “Adaptive restart for accelerated gradient schemes,” Found.Computational Math., vol. 15, no. 3, 715–32, Jun. 2015.
29 / 29
Approximate box blur
In practice, Bi � bi , so βi is usually a circle.The continuous-space PSF of a voxel slice at depth z is
∫ x+4x
2
x−4x2
∫ y+4y
2
y−4y
2
βi (s, t; x , y , z) dydx
≈∫ x+
4x2
x−4x2
∫ y+4y
2
y−4y
2
circ(s, t; cµL
i (x , y) , bi)dydx
=
∫ 4x2
−4x2
∫ 4y
2
−4y
2
circ(s, t; cµL
i (x + δx , y + δy ) , bi)dδydδx
=
∫∫circ(s − δs , t − δt ; cµL
i (x , y) , bi)· rect(δs , δt ;ws ,wt) dδtdδs .
Back
29 / 29
Infinitesimal vs finite-sized sensor pixel: point objectInfinitesimal voxel/pixel
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
4.8x 10
−3 Infinitesimal voxel
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
4.8x 10
−3
Infinitesimal voxel/pixel
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
4.8x 10
−3 Infinitesimal voxel
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
4.8x 10
−3
(For an infinitesimal voxel)29 / 29
Infinitesimal vs finite-sized sensor pixel: sphere object
Infinitesimal voxel/pixel
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
0.285
Infinitesimal voxel
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
0.285
Finite−sized voxel/pixel
s [mm]
t [m
m]
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
0
0.285
Infinitesimal voxel/pixel
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
0.285
Infinitesimal voxel
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
0.285
Finite−sized voxel/pixel
s [mm]
t [m
m]
−0.2 −0.1 0 0.1 0.2
−0.2
−0.1
0
0.1
0.2
0
0.285
point voxel point voxel finite voxelpoint sensor finite sensor finite sensor
29 / 29