v2009.01.01 - Convex Optimization
v2009.01.01 - Convex Optimization v2009.01.01 - Convex Optimization
328 CHAPTER 4. SEMIDEFINITE PROGRAMMING image/(reconstruction)error ratio 4.43 barrier that seems impenetrable by the total-variation objective. Total-variation minimization has met with moderate success, in retrospect, only because some medical images are moderately piecewise-constant signals. One simply hopes a reconstruction, that is in some sense equal to a known subset of samples and whose gradient is most sparse, is that unique image we seek. The total-variation objective, operating on an image, is expressible as norm of a linear transformation (765). It is natural to ask whether there exist other sparsifying transforms that might break the real-life 30dB barrier (@20% 256×256 data) in MRI. There has been much research into application of wavelets, discrete cosine transform (DCT), randomized orthogonal bases, splines, etcetera, but with suspiciously little focus on objective measures like image/error or illustration of difference images; the predominant basis of comparison instead being subjectively visual (Duensing & Huang, ISMRM Toronto 2008). 4.44 Despite choice of transform, there seems yet to have been a breakthrough of the 30dB barrier. Application of compressed sensing to MRI, therefore, remains fertile in 2008 for continued research. Lagrangian form of compressed sensing in imaging We now repeat Candès’ image reconstruction experiment from 2004 [61] which led to discovery of sparse sampling theorems. But we achieve perfect reconstruction with an algorithm based on vanishing gradient of the compressed sensing problem’s Lagrangian; our contraction method (p.334) [308,IIIA] is computationally 10× faster than contemporary methods because matrix multiplications are replaced by fast Fourier transforms 4.43 Noise considered here is due only to the reconstruction process itself; id est, noise in excess of that produced by the best reconstruction of an image from a complete set of samples in the sense of Shannon. At less than 30dB image/error, artifacts generally remain visible to the naked eye. We estimate about 50dB is required to eliminate noticeable distortion in a visual A/B comparison. 4.44 I have never calculated the PSNR of these reconstructed images [of Barbara]. −Jean-Luc Starck The sparsity of the image is the percentage of transform coefficients sufficient for diagnostic-quality reconstruction. Of course the term “diagnostic quality” is subjective. ...I have yet to see an “objective” measure of image quality. Difference images, in my experience, definitely do not tell the whole story. Often I would show people some of my results and get mixed responses, but when I add artificial Gaussian noise to an image, often people say that it looks better. −Michael Lustig
4.6. CARDINALITY AND RANK CONSTRAINT EXAMPLES 329 and number of constraints is cut in half by sampling symmetrically. Convex iteration for cardinality minimization is introduced which allows perfect reconstruction of a phantom at 4% subsampling rate; 50% Candès’ rate. By making neighboring pixel selection variable, we show image-gradient sparsity of the Shepp-Logan phantom to be 1.9% ; 33% lower than previously reported. We demonstrate application of image-gradient sparsification to the n×n=256×256 Shepp-Logan phantom, simulating ideal acquisition of MRI data by radial sampling in the Fourier domain (Figure 87). 4.45 Define a Nyquist-centric discrete Fourier transform (DFT) matrix ⎡ ⎤ 1 1 1 1 · · · 1 1 e −j2π/n e −j4π/n e −j6π/n · · · e −j(n−1)2π/n F = ∆ 1 e −j4π/n e −j8π/n e −j12π/n · · · e −j(n−1)4π/n 1 ⎢ 1 e −j6π/n e −j12π/n e −j18π/n · · · e −j(n−1)6π/n √ ∈ C n×n ⎥ n ⎣ . . . . ... . ⎦ 1 e −j(n−1)2π/n e −j(n−1)4π/n e −j(n−1)6π/n · · · e −j(n−1)2 2π/n (747) a symmetric (nonHermitian) unitary matrix characterized F = F T F −1 = F H (748) Denoting an unknown image U ∈ R n×n , its two-dimensional discrete Fourier transform F is F(U) ∆ = F UF (749) hence the inverse discrete transform U = F H F(U)F H (750) FromA.1.1 no.25 we have a vectorized two-dimensional DFT via Kronecker product ⊗ vec F(U) ∆ = (F ⊗F ) vec U (751) and from (750) its inverse [140, p.24] vec U = (F H ⊗F H )(F ⊗F ) vec U = (F H F ⊗ F H F ) vec U (752) 4.45 k-space is conventional acquisition terminology indicating domain of the continuous raw data provided by an MRI machine. An image is reconstructed by inverse discrete Fourier transform of that data sampled on a Cartesian grid in two dimensions.
- Page 277 and 278: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 279 and 280: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 281 and 282: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 283 and 284: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 285 and 286: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 287 and 288: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 289 and 290: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 291 and 292: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 293 and 294: 4.4. RANK-CONSTRAINED SEMIDEFINITE
- Page 295 and 296: 4.5. CONSTRAINING CARDINALITY 295 m
- Page 297 and 298: 4.5. CONSTRAINING CARDINALITY 297 m
- Page 299 and 300: 4.5. CONSTRAINING CARDINALITY 299 a
- Page 301 and 302: 4.5. CONSTRAINING CARDINALITY 301 f
- Page 303 and 304: 4.5. CONSTRAINING CARDINALITY 303 n
- Page 305 and 306: 4.5. CONSTRAINING CARDINALITY 305 W
- Page 307 and 308: 4.5. CONSTRAINING CARDINALITY 307 t
- Page 309 and 310: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 311 and 312: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 313 and 314: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 315 and 316: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 317 and 318: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 319 and 320: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 321 and 322: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 323 and 324: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 325 and 326: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 327: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 331 and 332: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 333 and 334: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 335 and 336: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 337 and 338: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 339 and 340: 4.6. CARDINALITY AND RANK CONSTRAIN
- Page 341 and 342: 4.7. CONVEX ITERATION RANK-1 341 fi
- Page 343 and 344: 4.7. CONVEX ITERATION RANK-1 343 Gi
- Page 345 and 346: Chapter 5 Euclidean Distance Matrix
- Page 347 and 348: 5.2. FIRST METRIC PROPERTIES 347 co
- Page 349 and 350: 5.3. ∃ FIFTH EUCLIDEAN METRIC PRO
- Page 351 and 352: 5.3. ∃ FIFTH EUCLIDEAN METRIC PRO
- Page 353 and 354: 5.4. EDM DEFINITION 353 The collect
- Page 355 and 356: 5.4. EDM DEFINITION 355 5.4.2 Gram-
- Page 357 and 358: 5.4. EDM DEFINITION 357 D ∈ EDM N
- Page 359 and 360: 5.4. EDM DEFINITION 359 5.4.2.2.1 E
- Page 361 and 362: 5.4. EDM DEFINITION 361 ten affine
- Page 363 and 364: 5.4. EDM DEFINITION 363 spheres: Th
- Page 365 and 366: 5.4. EDM DEFINITION 365 By eliminat
- Page 367 and 368: 5.4. EDM DEFINITION 367 where Φ ij
- Page 369 and 370: 5.4. EDM DEFINITION 369 5.4.2.2.6 D
- Page 371 and 372: 5.4. EDM DEFINITION 371 10 5 ˇx 4
- Page 373 and 374: 5.4. EDM DEFINITION 373 corrected b
- Page 375 and 376: 5.4. EDM DEFINITION 375 by translat
- Page 377 and 378: 5.4. EDM DEFINITION 377 Crippen & H
328 CHAPTER 4. SEMIDEFINITE PROGRAMMING<br />
image/(reconstruction)error ratio 4.43 barrier that seems impenetrable by<br />
the total-variation objective. Total-variation minimization has met with<br />
moderate success, in retrospect, only because some medical images are<br />
moderately piecewise-constant signals. One simply hopes a reconstruction,<br />
that is in some sense equal to a known subset of samples and whose gradient<br />
is most sparse, is that unique image we seek.<br />
The total-variation objective, operating on an image, is expressible<br />
as norm of a linear transformation (765). It is natural to ask<br />
whether there exist other sparsifying transforms that might break the<br />
real-life 30dB barrier (@20% 256×256 data) in MRI. There has been much<br />
research into application of wavelets, discrete cosine transform (DCT),<br />
randomized orthogonal bases, splines, etcetera, but with suspiciously little<br />
focus on objective measures like image/error or illustration of difference<br />
images; the predominant basis of comparison instead being subjectively<br />
visual (Duensing & Huang, ISMRM Toronto 2008). 4.44 Despite choice of<br />
transform, there seems yet to have been a breakthrough of the 30dB barrier.<br />
Application of compressed sensing to MRI, therefore, remains fertile in 2008<br />
for continued research.<br />
Lagrangian form of compressed sensing in imaging<br />
We now repeat Candès’ image reconstruction experiment from 2004 [61]<br />
which led to discovery of sparse sampling theorems. But we achieve<br />
perfect reconstruction with an algorithm based on vanishing gradient of the<br />
compressed sensing problem’s Lagrangian; our contraction method (p.334)<br />
[308,IIIA] is computationally 10× faster than contemporary methods<br />
because matrix multiplications are replaced by fast Fourier transforms<br />
4.43 Noise considered here is due only to the reconstruction process itself; id est, noise<br />
in excess of that produced by the best reconstruction of an image from a complete set<br />
of samples in the sense of Shannon. At less than 30dB image/error, artifacts generally<br />
remain visible to the naked eye. We estimate about 50dB is required to eliminate noticeable<br />
distortion in a visual A/B comparison.<br />
4.44 I have never calculated the PSNR of these reconstructed images [of Barbara].<br />
−Jean-Luc Starck<br />
The sparsity of the image is the percentage of transform coefficients sufficient for<br />
diagnostic-quality reconstruction. Of course the term “diagnostic quality” is subjective.<br />
...I have yet to see an “objective” measure of image quality. Difference images, in my<br />
experience, definitely do not tell the whole story. Often I would show people some of my<br />
results and get mixed responses, but when I add artificial Gaussian noise to an image,<br />
often people say that it looks better.<br />
−Michael Lustig