computational radar imaging
TRANSCRIPT
![Page 1: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/1.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Computational Radar Imaging
Müjdat Çetin
Associate Professor, Department of Electrical and Computer EngineeringInterim Director, Goergen Institute for Data Science
University of Rochester, Rochester, NY
Contributors: Burak Alver, Ammar Saleem, Özben Önhon, Sadegh Samadi, M. Javad Hasankhan, Abdurrahim Soğanlı, Emre Güven, Alper Güngör, Ivana Stojanovic,
Kush Varshney, W. Clem Karl, Randy Moses, Lee Potter, Alan S. Willsky.
![Page 2: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/2.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Radar Imaging Basics
•All-weather•Day and night operation•Superposition of response from scatterers – tomographic measurements
•Synthetic aperture radar (SAR)•Computational imaging problem: Obtain a spatial map of reflectivity from radar returns
![Page 3: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/3.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Outline
• Sparsity-driven radar imaging– Point-enhanced and region-enhanced imaging
– ADMM & proximal operators in the case of complex-valued fields
– Wide-angle imaging and anisotropy characterization
– Model errors and autofocusing
– Moving-object imaging
• Machine learning for radar imaging– Dictionary learning
– Deep learning-based priors
![Page 4: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/4.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Initial motivation for our work (circa 1999*)
• Accurate localization of dominant scatterers
– Limited resolution
– Clutter and artifact energy
Some challenges for automatic decision-making from SAR images:
• Region separability
– Speckle
– Object boundaries
• Limited or sparse apertures
* This slide was found in an archaeological excavation site and is believed to be ~2000 deep-learning-years old.
![Page 5: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/5.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
How I got attracted to sparsity…
![Page 6: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/6.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
SAR Ground-plane Geometry
• Scalar 2-D complex-valued reflectivity field
• Transmitted chirp signal:
• Received, demodulated return from circular patch:
![Page 7: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/7.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
SAR Observation Model
• Observations are related to projections of the field:
• SAR observations are band-limited slices from the 2-D Fourier transform of the reflectivity field:
• Discrete tomographic
SAR observation model:(combining all measurements)
Observed data
SAR Forward Model Unknown field
Noise
![Page 8: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/8.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Conventional Image Formation
• Given SAR returns, create an estimate of the reflectivity field f
Polar format algorithm:
• Each pulse gives slice of 2-D Fourier transform of field
• Polar to rectangular resampling
• 2-D inverse DFT
Support of observed data in the spatial frequency domain
Sample conventional image (reflectivity magnitudes)
![Page 9: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/9.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Sparsity-Driven Radar Imaging – basic version
• Bayesian interpretation: MAP estimation problem with heavy-tailed priors
• Complex-valued data and image
• Magnitude of complex-valued field admits sparse representation
• No informative prior on reflectivity phase
• Typical choices for L: – identity (point-enhanced imaging)
– gradient (region-enhanced imaging)
• Here focus on ADMM-based solutions
(f | y) (y | f) (f)p p p
![Page 10: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/10.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Setting up with a generic regularizer
• Discretized SAR observation model:
𝐲 = 𝐀𝐟 + 𝐧
• Retrieve 𝐟 using a regularized cost function:
𝐟 = argmin𝐟
{𝔇 𝐟 + 𝜆ℜ(|𝐟|)}
where 𝔇 𝐟 = 𝐲 − 𝐀𝐟 22 and ℜ(|𝐟|) is the regularizer
• Will describe two versions of ADMM
![Page 11: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/11.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Taking into account the complex-valued nature
• Rewrite 𝐟 = 𝚯 𝐟 where is 𝚯 a diagonal matrix containing the phase of 𝐟 in the form 𝑒𝑗𝜙(𝐟)
• Cost function becomes:
𝐟 , 𝚯 = argmin𝐟 ,𝚯
𝐲 − 𝐀𝚯 𝐟 22 + 𝜆ℜ( 𝐟 )
Reflectivity Magnitudes
Reflectivity Phases
![Page 12: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/12.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Variable Splitting and ADMM
• Introduce an auxiliary variable with a constraint:
𝐟 , 𝚯, 𝐡 = argmin𝐟 ,𝚯,𝐡
𝐲 − 𝐀𝚯 𝐟 22 + 𝜆ℜ 𝐡
𝑠. 𝑡. 𝐟 − 𝐡 = 0
• Augmented Lagrangian (in scaled form):
𝐟 , 𝚯, 𝐡, 𝐮 = argmin𝐟 ,𝚯,𝐡,𝐮
𝐲 − 𝐀𝚯 𝐟 22 + 𝜆ℜ 𝐡
+𝜌
2𝐟 − 𝐡 + 𝐮 2
2 +𝜌
2𝐮 2
2
![Page 13: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/13.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Variables involved - details
• Let 𝜽 ∈ ℂ𝑁×1 be a vector containing the diagonal elements of the phase matrix 𝚯
• Invoke the constraint that the magnitudes of the elements of 𝜽 should be 1, since they contain phases in the form 𝑒𝑗𝜙(𝐟)
• Let 𝐁 be a matrix whose diagonal elements contain the reflectivity magnitudes
• Let 𝐟 = 𝐡 − 𝐮 and 𝐡 = 𝐟 + 𝐮
![Page 14: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/14.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
ADMM for Complex-valued Imaging
• Each iteration of the ADMM algorithm performs the followingsteps enabling the use of a Plug-and-Play (PnP) prior approach:
𝜽 = argmin𝜽
𝐲 − 𝐀𝐁𝜽 22 + 𝜆𝛉 𝑖=1
𝑁 ( 𝜽𝑖 − 1)2
𝐟 = argmin𝐟
𝐲 − 𝐀𝚯 𝐟 22 +
𝝆
2𝐟 − 𝐟
2
2
𝐡 = argmin𝐡
𝜆ℜ 𝐡 +𝝆
2 𝐡 − 𝐡
2
2
𝐮 = 𝐮 + 𝐟 − 𝐡
where 𝜆𝛉 is a hyperparameter
Data
Prior
Data
All real-valued
![Page 15: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/15.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Alternative way to set up ADMM
• Introduce an auxiliary variable with a constraint:
𝐟, 𝐡 = argmin𝐟,𝐡
𝐲 − 𝐀𝐟 22 + 𝜆ℜ |𝐡|
𝑠. 𝑡. 𝐟 − 𝐡 = 0
• Augmented Lagrangian (in scaled form):
𝐟, 𝐡, 𝐮 = argmin𝐟,𝐡,𝐮
𝐲 − 𝐀𝐟 22 + 𝜆ℜ |𝐡|
+𝜌
2𝐟 − 𝐡 + 𝐮 2
2 +𝜌
2𝐮 2
2
• Let 𝐟 = 𝐡 − 𝐮 and 𝐡 = 𝐟 + 𝐮
![Page 16: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/16.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
ADMM-II for Complex-valued Imaging
• Each iteration of the ADMM algorithm performs the followingsteps enabling the use of a Plug-and-Play (PnP) prior approach:
𝐟 = argmin𝐟
𝐲 − 𝐀𝐟 22 +
𝝆
2𝐟 − 𝐟
2
2
𝐡 = argmin𝐡
𝜆ℜ |𝐡| +𝝆
2 𝐡 − 𝐡
2
2
𝐮 = 𝐮 + 𝐟 − 𝐡
where 𝜆𝛉 is a hyperparameter
Prior
Data
Complex-valued with
regularization on magnitudes
![Page 17: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/17.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Proximal Mappings for Complex-Valued Problems
• Consider total variation (TV) regularization.
• Proximal mapping for standard TV can be computed by efficient algorithms
• We want to penalize TV of the magnitude of a complex-valued field
• Proximal mapping operator for TV-magnitude:
• Similarly, proximal mapping operator of a linear transformation (Wavelet trf., DCT,…) of the magnitude of the complex field:
![Page 18: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/18.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Proximal Mapping for lp–norm-based Regularizers
• Proximal mapping for lp-norm
• Can be computed through a series of iteratively reweighted soft thresholding operations:
![Page 19: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/19.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Sparsity-Driven SAR Imaging Results
Point-enhanced Region-enhanced Wavelet dictionary
![Page 20: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/20.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Conventional Sparsity-driven, ADMM-based
Sparsity-Driven SAR Imaging Results
![Page 21: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/21.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Non-Traditional Apertures: Wide-Angle Data
• Irregular PSF
• Anisotropic scattering • Conventional imaging produces
inaccurate reflectivities
• Conventional imaging does not characterize aspect dependence
![Page 22: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/22.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Anisotropic Scattering in SAR
• Much reflection from one angle
![Page 23: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/23.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Anisotropic Scattering in SAR
• Little reflection from another angle
![Page 24: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/24.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Composite Image Formation• Many scatterers won’t persist over wide angles
• Form a composite wide-angle image from narrow-angle subaperture images:
Subaperture index
Pixel indices
k-th subaperture image
Composite image
• Preservation of anisotropic scatterers with short persistence
• Partial characterization of aspect dependence
![Page 25: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/25.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
• Angular range = 110°
• Composite images
Reference image Conventional Sparsity-driven
25
Wide-angle SAR Imaging
![Page 26: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/26.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
• 70% of the band available
Reference image Conventional Sparsity-driven
Wide-Angle SAR Imaging with Frequency-Band Omissions
angle
freq
blocked
![Page 27: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/27.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Visualization of aspect dependencein wide-angle imaging
![Page 28: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/28.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Joint Imaging and Anisotropy Characterization
• Model for data collected from P isotropicscatterers:
– Image formation: Recovering f(x,y)
• When we consider anisotropy :
– Joint imaging and anisotropy characterization:Recovering f(x,y,θ)
• The problem becomes much more underdetermined!
P
m
mmmm yxc
jyxfr1
sincos2
exp,,
P
m
mmmm yxc
jyxfr1
sincos2
exp,,,
![Page 29: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/29.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Modeling Angular Anisotropy• Model f(x,y,θ) by the dictionary {b1(θ), b2(θ), …, bM(θ)}
• Problem: determine unknown coefficients am,I
• bi(θ): atoms of the angular response dictionary
• Construct an overcomplete dictionary to allow sparse representation of typical angular response patterns.
P
m
M
i
mmmmiim yxc
jyxbar1 1
, sincos2
exp,,,
P
m
mmmm yxc
jyxfr1
sincos2
exp,,,
![Page 30: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/30.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Sample Result of an Analysis-based Formulation for Anisotropy Characterization
Aspect dependent scattering
behavior in indicated sub-area
x
y
![Page 31: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/31.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Joint Imaging and Model Error Correction: Sparsity-Driven Autofocus (SDA)
• SAR observation model may not be known perfectly, due to, e.g., uncertainties in platform location or atmospheric turbulence
• This leads to phase errors in observed data
• Proposed approach: joint optimization over the reflectivities and model parameters:
• Solution through coordinate descent or ADMM− Closed-form solution for the phase error estimation step
1
2
2)(),( ffAyf J
31
![Page 32: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/32.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Sparsity-Driven Autofocus (SDA)
Without phaseerrors
Conventional imaging Sparsity-driven imaging
Conventional imaging Sparsity-driven imaging Proposed SDA
1D Phase error uniform. dist. in
2,
2
32
Random, uniform phaseerror
![Page 33: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/33.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Synthesis-based SDA
Reference conventional
image without
phase errors
Proposed approach
with phase errors
Conventional image
with phase errors
33
Wavelet dictionary
![Page 34: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/34.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Moving-Target Imaging
• SAR platform position uncertainties cause space-invariant defocusing of the reconstructed image, i.e., the amount of defocusing is the same for all points in the scene.
• Motion of a target in the scene can also be modeled as a phase error over the phase history data corresponding to a stationary scene.
• Moving targets in the scene cause artifacts including defocusing around the spatial neighborhood of the target in the scene.
Space-variant defocusing
Need to keep an account of the contributions from each spatial location to the phase error at each aperture position
![Page 35: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/35.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Sparsity-Driven Moving-Target Imaging
iits
ffAyfJff
1)(..
)(minarg),(minarg1211
2
2,,
1
Tmjmjmj
mIeee
)()()(,.......,, 21
M
.
.
.
2
1
• Involves sparsity constraints both on the reflectivity field and on the motion field.
• Have also developed a more efficient and potentially robust version based on constructing ROIs and performing space-invariant focusing within them.
![Page 36: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/36.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Moving-Target ImagingProposed methodSparsity-driven imagingConventional imaging
36
![Page 37: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/37.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Moving Target Imaging by Low-rank and Sparse Decomposition
![Page 38: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/38.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Dictionary Learning for Sparsity-Driven SAR Imaging
Training Images
K-SVD
Learned
dictionary
Conventional Learning-based
![Page 39: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/39.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Deep Learning-based Priors for SAR Imaging
• A new SAR image reconstruction framework utilizing Plug-and-Play (PnP) priors– Optimization-based reconstruction - regularized
inversion,
– Decoupling the data and the prior through ADMM
– Deep learning-based prior
![Page 40: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/40.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Variable Splitting and ADMM
• Each iteration of the ADMM algorithm performs the followingsteps enabling the use of a Plug-and-Play (PnP) prior approach:
𝜽 = argmin𝜽
𝐲 − 𝐀𝐁𝜽 22 + 𝜆𝛉 𝑖=1
𝑁 ( 𝜽𝑖 − 1)2
𝐟 = argmin𝐟
𝐲 − 𝐀𝚯 𝐟 22 +
𝝆
2𝐟 − 𝐟
2
2
𝐡 = argmin𝐡
𝜆ℜ 𝐡 +𝝆
2 𝐡 − 𝐡
2
2
𝐮 = 𝐮 + 𝐟 − 𝐡
where 𝜆𝛉 is a hyperparameter
Data
Prior
Data
![Page 41: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/41.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Convolutional Neural Network (CNN)-based Prior
• Architecture: modified version of the network used in [1]
• 20 convolutional modules
• First and even numbered modules: 64 3×3 filters withpadding 1, stride 1
• Remaining modules: 64 5×5 filters with padding 2, stride 1
• Each module has batch normalization and ReLU layers
[1] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deepCNN for image denoising,” IEEE Transactions on Image Processing, 26(7):3142-3155, 2017.
![Page 42: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/42.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Training the CNN - Synthetic ScenesUsing 64×64 ground truth images for CNN training:
• Add random phase to training images and obtain the phasehistories.
• Apply complex-valued additive noise to the phase histories.– Magnitude uniformly distributed over [0, 𝜎𝑦], where 𝜎𝑦 is the standard
deviation of the magnitude of the phase history data
– Phase uniformly distributed over [−𝜋, 𝜋]
• Perform conventional reconstructions.
• Extract 16×16 overlapping patches from these conventionalimages and their corresponding ground truths, to constructinput-output pairs.
• Augment the pairs of images through rotation by[90°, 180°, 270°].
• Train the network using these augmented pairs of images, with image reconstructed from noisy data as input and ground truth image as output.
![Page 43: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/43.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Synthetic Data Experiments --Training Set
![Page 44: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/44.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Test Set
• 5 different phase history data availability levels:
100%, 87.89%, 76.56%, 56.25%, 25%
• 2 different noise levels (𝜎𝑛 ∈ 0.1, 1 𝜎𝑦)
• Rectangular band-limitation for data reduction
![Page 45: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/45.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 7, 𝜎𝑛 = 0.1𝜎𝑦 and full data
![Page 46: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/46.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 7, 𝜎𝑛 = 0.1𝜎𝑦 and 87.89% data
![Page 47: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/47.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 7, 𝜎𝑛 = 0.1𝜎𝑦 and 76.56% data
![Page 48: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/48.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 7, 𝜎𝑛 = 0.1𝜎𝑦 and 56.25% data
![Page 49: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/49.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 7, 𝜎𝑛 = 0.1𝜎𝑦 and 25% data
![Page 50: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/50.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Quantitative Results
![Page 51: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/51.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Preliminary Results on Real Scenes from TerraSAR-X
• Training based on the Netherlands Rotterdam Harbor StaringSpotlight SAR image (1041 × 1830)
– Split into 448 non-overlapping 64 × 64 ‘‘windows’’
– 1075648 overlapping 16 × 16 patches extracted from windows
– Patches augmented with rotations of 90°, 180°, 270°
• Test set: 751 selected windows extracted from the Panama High Resolution Spotlight SAR image (2375 × 3375)
![Page 52: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/52.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 608, 𝜎𝑛 = 𝜎𝑦, full data
![Page 53: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/53.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 608, 𝜎𝑛 = 𝜎𝑦, 87.89% data
![Page 54: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/54.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 608, 𝜎𝑛 = 𝜎𝑦, 76.56% data
![Page 55: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/55.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 608, 𝜎𝑛 = 𝜎𝑦, 56.25% data
![Page 56: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/56.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Qualitative Results: Image 608, 𝜎𝑛 = 𝜎𝑦, 25% data
![Page 57: Computational Radar Imaging](https://reader030.vdocuments.us/reader030/viewer/2022012519/619451555a8b092d4f3b1ddb/html5/thumbnails/57.jpg)
Müjdat Çetin IMA Computational Imaging Workshop, October 14-18, 2019
Conclusion• A line of inquiry that involves:
– Radar sensing
– Computational imaging
– Signal representation, compressed sensing
– Machine learning
• Sparsity is a useful asset for radar imaging especially in nonconventional
data collection scenarios (e.g., when the data are sparse, irregular, limited)
• Deep learning methods may have the potential to learn complicated spatial patterns and enable their incorporation as priors into computational radar imaging
• Radar imaging offers a rich set of inference problems that can benefit from machine learning
– Reflectivity field, anisotropy, model (parameters), object motion
– Nonlinear case, multiple scattering, multi-static imaging, passive imaging
• Several challenging yet potentially rewarding themes for the near future:
– Confidence/uncertainty characterization, posterior sampling
– Machine learning for computational imaging and/or decision making