compressive sensing -- a 25 minute tour
TRANSCRIPT
![Page 1: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/1.jpg)
Compressive Sensing – A 25 Minute Tour
Emmanuel Candes
First EU-US Frontiers of Engineering Symposium, Cambridge, September 2010
![Page 2: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/2.jpg)
Dedication
Dennis Healy, 1957 – 2009
![Page 3: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/3.jpg)
Compressive sensing today
Enormous field spanning
mathematics
applied mathematics
computer science
information theory
signal processing
circuit design
optical engineering
biomedical imaging
...
Many contributors; e. g. R. Baraniuk
Compressed sensingtoday
http://nuit-blanche.blogspot.com/
http://dsp.rice.edu/cs
Our focus: short and biased overview
![Page 4: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/4.jpg)
A contemporary paradox
Raw: 15MB JPEG: 150KB
Massive data acquisition
Most of the data is redundant and can be thrown away
Seems enormously wasteful
One can regard the possibility of digital compression as a failure ofsensor design. If it is possible to compress measured data, one mightargue that too many measurements were taken.
David Brady
![Page 5: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/5.jpg)
A contemporary paradox
Raw: 15MB JPEG: 150KB
Massive data acquisition
Most of the data is redundant and can be thrown away
Seems enormously wasteful
One can regard the possibility of digital compression as a failure ofsensor design. If it is possible to compress measured data, one mightargue that too many measurements were taken.
David Brady
![Page 6: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/6.jpg)
Going against a long established tradition?
Acquire/Sample (A-to-D converter, digital camera)
Compress (signal dependent, nonlinear)
compress transmit/store
receive decompress
sample
sparse
wavelettransform
!
N
!
N >> M
!
M
!
M
!
N
Fundamental question
Can we directly acquire just the useful part of the signal?
![Page 7: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/7.jpg)
What Is Compressive Sensing?
![Page 8: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/8.jpg)
In a nutshell...
Can obtain super-resolved signals from just a few sensors
Sensing is nonadaptive: no effort to understand the signal
Simple acquisition process followed by numerical optimization
First papers
Candes, Romberg and Tao, 2006
Candes and Tao, 2006
Donoho, 2006
By now, very rich mathematical theory
![Page 9: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/9.jpg)
Sparsity: wavelets and images
1 megapixel image
wavelet coefficients
0 2 4 6 8 10 12
x 105
−4000
−2000
0
2000
4000
6000
8000
10000
12000
14000
16000
⇓
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7
x 104
−3000
−2000
−1000
0
1000
2000
3000
zoom in
Images have a sparse representation in the wavelet domain
![Page 10: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/10.jpg)
Sparsity: wavelets and images
1 megapixel image
wavelet coefficients
0 2 4 6 8 10 12
x 105
−4000
−2000
0
2000
4000
6000
8000
10000
12000
14000
16000
⇓
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7
x 104
−3000
−2000
−1000
0
1000
2000
3000
zoom in
Images have a sparse representation in the wavelet domain
![Page 11: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/11.jpg)
Implication of sparsity: image “compression”
1 Compute 1,000,000 wavelet coefficients of mega-pixel image
2 Set to zero all but the 25,000 largest coefficients
3 Invert the wavelet transform
original image
after zeroing out smallest coefficients
This principle underlies modern lossy coders (sound, still-picture, video)
![Page 12: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/12.jpg)
Implication of sparsity: image “compression”
1 Compute 1,000,000 wavelet coefficients of mega-pixel image
2 Set to zero all but the 25,000 largest coefficients
3 Invert the wavelet transform
original image after zeroing out smallest coefficients
This principle underlies modern lossy coders (sound, still-picture, video)
![Page 13: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/13.jpg)
Idealized sampling
x: signal coefficients in our convenient representation
collect information by measuring largest components of x
0 5 10 15 20 25 30 35 40 45 50!1
!0.8
!0.6
!0.4
!0.2
0
0.2
0.4
0.6
0.8
1
! " #! #" $! $" %! %" &! &" "!!#
!!'(
!!')
!!'&
!!'$
!
!'$
!'&
!')
!'(
#
sparse x nearly sparse x
What if these positions are not known in advance?
what should we measure?
how should we reconstruct?
![Page 14: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/14.jpg)
Idealized sampling
x: signal coefficients in our convenient representation
collect information by measuring largest components of x
0 5 10 15 20 25 30 35 40 45 50!1
!0.8
!0.6
!0.4
!0.2
0
0.2
0.4
0.6
0.8
1
! " #! #" $! $" %! %" &! &" "!!#
!!'(
!!')
!!'&
!!'$
!
!'$
!'&
!')
!'(
#
sparse x nearly sparse x
What if these positions are not known in advance?
what should we measure?
how should we reconstruct?
![Page 15: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/15.jpg)
Incoherent/random sensing
y = 〈ak, x〉, k = 1, . . . ,m
Want sensing waveforms as spread out/“incoherent” as possible
Span of {ak} should be as random as possible (general orientation)
aki.i.d.∼ F
E aka∗k = I and ak spread outak i.i.d. N (0, 1) (white noise)ak i.i.d. ±1ak = exp(i2πωkt) with i.i.d. frequencies ωk
...
0 20 40 60 80 100 120!2.5
!2
!1.5
!1
!0.5
0
0.5
1
1.5
2
2.5
0 20 40 60 80 100 120!2.5
!2
!1.5
!1
!0.5
0
0.5
1
1.5
2
2.5
0 100 200 300 400 500 600 700 800 900 1000!2
!1.5
!1
!0.5
0
0.5
1
1.5
2
random waveform N(0, 1) random waveform ±1 random sinusoid
![Page 16: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/16.jpg)
Incoherence
Motivation: Sampling a Sparse Vector
concentrated vector incoherent measurements
• Signal is local, measurements are global
• Each measurement picks up a little information about each component
• Triangulate significant components from measurements
• Formalization: Relies on uncertainty principles between sparsity basis andmeasurement system
Signal is local, measurements are global
Each measurement picks up a little information about each component
![Page 17: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/17.jpg)
Example of foundational result
Classical viewpoint
Measure everything (all the pixels, all the coefficients)
Keep d largest coefficients: distortion is ‖x− xd‖
Compressed sensing viewpoint
Take m random measurements: yk = 〈x, ak〉Reconstruct by linear programming: (‖x‖`1 =
∑i |xi|)
x? = arg min ‖x‖`1 subject to yk = 〈x, ak〉, k = 1, . . . ,m
Among all the objects consistent with data, pick min `1
Same performance with about m = d log n/d (sketch)
‖x? − x‖`2 ≤ ‖x− xd‖`2
![Page 18: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/18.jpg)
Example of foundational result
Classical viewpoint
Measure everything (all the pixels, all the coefficients)
Keep d largest coefficients: distortion is ‖x− xd‖
Compressed sensing viewpoint
Take m random measurements: yk = 〈x, ak〉Reconstruct by linear programming: (‖x‖`1 =
∑i |xi|)
x? = arg min ‖x‖`1 subject to yk = 〈x, ak〉, k = 1, . . . ,m
Among all the objects consistent with data, pick min `1
Same performance with about m = d log n/d (sketch)
‖x? − x‖`2 ≤ ‖x− xd‖`2
![Page 19: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/19.jpg)
Example of foundational result
Classical viewpoint
Measure everything (all the pixels, all the coefficients)
Keep d largest coefficients: distortion is ‖x− xd‖
Compressed sensing viewpoint
Take m random measurements: yk = 〈x, ak〉Reconstruct by linear programming: (‖x‖`1 =
∑i |xi|)
x? = arg min ‖x‖`1 subject to yk = 〈x, ak〉, k = 1, . . . ,m
Among all the objects consistent with data, pick min `1
Same performance with about m = d log n/d (sketch)
‖x? − x‖`2 ≤ ‖x− xd‖`2
![Page 20: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/20.jpg)
Example
Take 96K incoherent measurements of “compressed” image
Compressed image is perfectly sparse (25K nonzero wavelet coeffs)
Solve `1
original (25k wavelets)
perfect recovery
![Page 21: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/21.jpg)
Example
Take 96K incoherent measurements of “compressed” image
Compressed image is perfectly sparse (25K nonzero wavelet coeffs)
Solve `1
original (25k wavelets) perfect recovery
![Page 22: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/22.jpg)
What is compressive sensing?
Possibility of compressed data acquisition protocols which directlyacquire just the important information
Incoherent/random measurements → compressed description
Simultaneous signal acquisition and compression!
All we need is to decompress...
Three surprises
Sensing is ultra efficient and nonadaptive
Recovery is possible by tractable optimization
Sensing/recovery is robust to noise (and other imperfections)
![Page 23: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/23.jpg)
What is compressive sensing?
Possibility of compressed data acquisition protocols which directlyacquire just the important information
Incoherent/random measurements → compressed description
Simultaneous signal acquisition and compression!
All we need is to decompress...
Three surprises
Sensing is ultra efficient and nonadaptive
Recovery is possible by tractable optimization
Sensing/recovery is robust to noise (and other imperfections)
![Page 24: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/24.jpg)
Applications and Opportunities
![Page 25: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/25.jpg)
Three potentially impacted areas
Analog-to-digital conversion
Optical systems
Magnetic Resonance Imaging
![Page 26: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/26.jpg)
Time or space sampling
Analog-to-digital converters, receivers, ...
Space: cameras, medical imaging devices, ...
Nyquist/Shannon foundation: must sample at twice the highest frequency
![Page 27: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/27.jpg)
Sampling of ultra wideband radio frequency signals
![Page 28: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/28.jpg)
Hardware brick wall
Signals are wider and wider band
“Moore’s law:” factor 2 improvement every 6 to 8 years
Extremely fast/high-resolution samplers are decades away
Way out? Analog-to-information (DARPA)
![Page 29: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/29.jpg)
Hardware brick wall
Signals are wider and wider band
“Moore’s law:” factor 2 improvement every 6 to 8 years
Extremely fast/high-resolution samplers are decades away
Way out? Analog-to-information (DARPA)
![Page 30: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/30.jpg)
What have we learned?
If ‘information bandwidth’ less than total bandwidth, then should be able to
sample below Nyquist without information loss
recover missing samples by convex optimization
f (ω)
ω t
f(t)
![Page 31: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/31.jpg)
New analog-to-digital converters
Joint with Caltech and Northrop Grumman
NUS (InP) RMPI (CMOS) RMPI (InP)
Manufactured three chips
Nonuniform sampler (InP)
Random modulator pre-integration
4 channels (InP)
8 channels (CMOS)
![Page 32: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/32.jpg)
Random modulator pre-integrator (RMPI)
Joint with Becker, Emami and Yoo (Caltech)
∫kT1,-1,1,1,-1,...
yk,1
∫kT-1,-1,-1,1,-1,...
yk,2
∫kT-1,1,-1,1,1,...
yk,P
f(t)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1!2
!1.5
!1
!0.5
0
0.5
1
1.5
2
RMPI Basic architecture random modulation
Nyquist rate 5GHz
8 channels at 50MHz → 400MHz
12.5× undersampled
![Page 33: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/33.jpg)
RMPI chip v.2
Process:IBM 90nm CMOS9SF 06 02 00 LB
Power Consumption: 830mW
Bandwidth: 2.5GHz
Dynamic Range: 50dB
Sampling rate: 400 MS/s
Die Area: 2mm× 2mm
![Page 34: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/34.jpg)
Pulse recovery from full system simulations
0 0.5 1 1.5 2 2.5−190
−180
−170
−160
−150
−140
−130
−120
−110
Frequency (GHz)
Pow
er/fr
eque
ncy
(dB
/Hz)
Frequency domain. Two overlapping pulses at 103.3 MHz and 2410.6 MHz carriersAmplitude 1: 3.2e−03, Amplitude 2: 1.0e−02 (10.0 dB dynamic range)
N=2048 (T=410 ns), jitter rms = 0.5 psEstimate pulse 1 to have frequency 2410.6 MHz (true value 2410.6 MHz)Estimate pulse 2 to have frequency 103.3 MHz (true value 103.3 MHz)
original
recovered
0 40 80 120 160 200 240 280 320 360 400−0.01
−0.005
0
0.005
0.01
0.015
time (ns)
Time domain., MSE=6.2⋅10−3
Rel. l2 error: 7.8⋅10−2; rel. l∞ error: 1.3⋅10−1; rel. l∞ freq error: 5.0⋅10−2
MSE on pulse envelopes: 5.0⋅10−3, 1.9⋅10−3
TOA estimate: 66.6 ns, Actual: 67.0 ns, Error: 0.4 ns Duration estimate: 203.4 ns, Actual: 199.8 ns, Error: 3.6 nsTOA estimate: 107.2 ns, Actual: 110.2 ns, Error: 3.0 ns Duration estimate: 206.2 ns, Actual: 200.0 ns, Error: 6.2 ns
demod of exact signal at 2411 MHzdemod of recovered signal at 2411 MHzdemod of exact signal at 103 MHzdemod of recovered signal at 103 MHzerror
![Page 35: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/35.jpg)
Pulse recovery from full system simulations
0 0.5 1 1.5 2 2.5−190
−180
−170
−160
−150
−140
−130
−120
−110
Frequency (GHz)
Pow
er/fr
eque
ncy
(dB
/Hz)
Frequency domain. Two overlapping pulses at 703.3 MHz and 2103.6 MHz carriersAmplitude 1: 1.0e−02, Amplitude 2: 1.0e−02 (0.0 dB dynamic range)
N=2048 (T=410 ns), jitter rms = 0.5 psEstimate pulse 1 to have frequency 2103.6 MHz (true value 2103.6 MHz)Estimate pulse 2 to have frequency 703.3 MHz (true value 703.3 MHz)
original
recovered
0 40 80 120 160 200 240 280 320 360 400−0.01
−0.005
0
0.005
0.01
0.015
time (ns)
Time domain., MSE=1.1⋅10−3
Rel. l2 error: 3.4⋅10−2; rel. l∞ error: 5.4⋅10−2; rel. l∞ freq error: 2.6⋅10−2
MSE on pulse envelopes: 1.5⋅10−3, 2.6⋅10−4
TOA estimate: 66.6 ns, Actual: 67.0 ns, Error: 0.4 ns Duration estimate: 200.8 ns, Actual: 199.8 ns, Error: 1.0 nsTOA estimate: 107.2 ns, Actual: 110.2 ns, Error: 3.0 ns Duration estimate: 204.2 ns, Actual: 200.0 ns, Error: 4.2 ns
demod of exact signal at 2104 MHzdemod of recovered signal at 2104 MHzdemod of exact signal at 703 MHzdemod of recovered signal at 703 MHzerror
![Page 36: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/36.jpg)
Real testing: NGC-Caltech A-to-I Receiver Test(Die at similar scale)
Caltech CMOS RMPI test hardware (RMPI, ADC, and FPGA/DAQ boards)
![Page 37: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/37.jpg)
Real results
Photo of RMPI AIC Breadboard
RMPI Chip ADC
ADC
A DC
AD C RMPI (InP) behaves as simulated
2.5GHz of bandwidth
50-60dB of dynamic range
∼ 3W of power consumption
Other ADC efforts
Eldar et al. (Technion)
Fudge et al. (L3 & Wisconsin)
Baraniuk et al. (Rice)
![Page 38: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/38.jpg)
Three potentially impacted areas
Analog-to-digital conversion
Optical systems
Magnetic Resonance Imaging
![Page 39: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/39.jpg)
What do we measure?
Direct sampling: analog/digital photography,mid 19th century
Indirect sampling: acquisition in atransformed domain, second half of 20thcentury; e.g. CT, MRI
Compressive sampling: acquisition in anincoherent domain
Design incoherent analog sensors ratherthan usual pixelsPay-off: need far fewer sensors
The first photograph?
CT scanner
![Page 40: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/40.jpg)
One pixel camera
Richard Baraniuk, Kevin Kelly, Yehia Massoud, Don JohnsonRice University, dsp.rice.edu/CS
MIT Tech review: top 10 emerging technologies for 2007
Other works: Brady et al., Freeman et al., Wagner et al., Coifman et al.
![Page 41: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/41.jpg)
“Single-Pixel” CS Camera
DMD DMD
single photon
detector
Kevin Kelly
Richard Baraniuk
Rice University
scene
can be exotic
• IR, UV, THz, PMT,
spectrometer, …
random
pattern on DMD array
(spatial light modulator)
raster scan IR CS IR hyperspectral data cube
color
target
![Page 42: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/42.jpg)
Three potentially impacted areas
Analog-to-digital conversion
Optical systems
Magnetic Resonance Imaging
![Page 43: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/43.jpg)
Fast Magnetic Resonance Imaging
Goal: sample less to speed up MR imaging process
Fully sampled 6 × undersampled 6 × undersampledclassical CS
Trzasko, Manduca, Borisch (Mayo Clinic)
![Page 44: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/44.jpg)
MR angiography
Fully sampled 6 × undersampled
Trzasko, Manduca, Borisch
![Page 45: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/45.jpg)
Fully sampled 6 × undersampled
Trzasko, Manduca, Borisch
![Page 46: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/46.jpg)
Compressive sensing in the news
Wired, March 2010
![Page 47: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/47.jpg)
l1-SPIRiT, 1st Pass T1 3D SPGR
Maximum Intensity Projection (MIP) PI
Slices l1-SPIRiT
Breath-hold post-gadolinium MRA in a 9 year old male with hypertension using 4X acceleration at 1.2 mm3 resolution. Left images (a, c) are with ARC and right (b, d) are with L1-SPIRIT compressed sensing. Note improved delineation of pancreas (big arrow), pancreatic duct (middle arrow), and diaphragm (small arrow) with l1-SPIRiT. Left gastric artery (arrowhead) emerges from the noise.
M. Lustig, Electrical Eng. Stanford & EECS UC Berkeley S. Vasanawala, Radiology Stanford
![Page 48: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/48.jpg)
l1-SPIRiT, T1 3D SPGR
Submillimeter near-isotropic resolution MRI in an 8-year-old male. Post-contrast T1 imaging with an acceleration of 4. Standard (a, c) and compressed sensing reconstruction (b, d) show improved delineation of the pancreatic duct (vertical arrow), bowel (horizontal arrow), and gallbladder wall (arrowhead) with L1-SPIRiT reconstruction, and equivalent definition of the portal vein (black arrow)
M. Lustig, Electrical Eng. Stanford & EECS UC Berkeley S. Vasanawala, Radiology Stanford
![Page 49: Compressive Sensing -- A 25 Minute Tour](https://reader030.vdocuments.net/reader030/viewer/2022012616/619e5571cc902b51b456b426/html5/thumbnails/49.jpg)
Summary
Ultra efficient acquisition protocol:
automatically translates analog data into already compressed digital form
Change the way engineers think about data acquisition
Already many applications
More applications to come