10.03.2015 Views

v2009.01.01 - Convex Optimization

v2009.01.01 - Convex Optimization

v2009.01.01 - Convex Optimization

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

4.6. CARDINALITY AND RANK CONSTRAINT EXAMPLES 327<br />

June 2004: Candès, Romberg, & Tao publish in cyberspace: 5481 complex samples<br />

(22 radial lines, ≈256 complex samples per) to reconstruct noiseless<br />

256×256-pixel Shepp-Logan phantom by 1-norm minimization of an<br />

image-gradient integral estimate called total variation; id est, 8.4%<br />

subsampling of 65536 data. [61,1.1] [60,3.2]<br />

March 2006: Chartrand discovers 4.40 that reconstruction of the Shepp-Logan<br />

phantom is possible with only 2521 complex samples (10 radial lines,<br />

Figure 87); 3.8% subsampled data input to a (nonconvex) 1 2 -norm<br />

total-variation minimization. [67,IIIA]<br />

April 2007: Trzasko & Manduca recover the phantom with same subsampling<br />

as Chartrand, but they realize an accurate 0-norm total-variation<br />

minimization algorithm that is 10× faster 4.41 than previous methods.<br />

[308] [309]<br />

Passage of these few years witnessed an algorithmic speedup and dramatic<br />

reduction in minimum number of samples required for perfect reconstruction<br />

of the noiseless Shepp-Logan phantom. But minimization of total variation<br />

is ideally suited to recovery of any piecewise-constant image, like a phantom,<br />

because the gradient of such images is highly sparse by design.<br />

There is no inherent characteristic of real-life MRI images that would<br />

make reasonable an expectation of sparse gradient. Sparsification of an<br />

image-gradient tends to preserve edges. Then minimization of total variation<br />

seeks an image having fewest edges. There is no deeper theoretical foundation<br />

than that. When applied to human brain scan or angiogram, with as<br />

much as 20% of 256×256 Fourier samples, we have observed 4.42 a 30dB<br />

4.40 I try using p = 1 2<br />

in place of p =1[-norm] to reconstruct the Shepp-Logan phantom,<br />

and am very surprised to see how well it works. This is by alternating gradient descent<br />

with projection onto the constraint. In hindsight, one aspect appears to have been crucial<br />

to its success: regularization of the cusp with a relatively large epsilon, then decreasing<br />

epsilon after convergence. At the time I did this solely to try to speed up convergence,<br />

based on experience with TV minimization algorithms (like lagged diffusivity). But it now<br />

seems to be what allows the algorithm to avoid local minima, as these are filled in by the<br />

larger epsilon.<br />

−Rick Chartrand<br />

4.41 Their Matlab code runs in a minute, on an Intel-based dual-processor 2006 vintage<br />

laptop computer, capable of recovery at 50dB image/error from 3.2% subsampled data.<br />

4.42 Work on real-life images was performed in conjunction with Christine Law at Lucas<br />

Center for Imaging, Stanford University.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!