10.03.2015 Views

sparse image representation via combined transforms - Convex ...

sparse image representation via combined transforms - Convex ...

sparse image representation via combined transforms - Convex ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

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

6.6. SUMMARY OF COMPUTATIONAL EXPERIMENTS 127<br />

6.6 Summary of Computational Experiments<br />

We ran our experiments on an SGI Power Challenger server with 196 MHz CPU. In the<br />

algorithm in Chapter 4, we start with γ =10 2 andstopwithγ =10 5 . Table 6.1 gives<br />

the number of LSQR iterations and the execution time of one iteration in each case. Each<br />

LSQR iteration includes one analysis transform and one synthesis transform of an <strong>image</strong>.<br />

Since the number of LSQR iterations is machine independent, it is a good criterion for<br />

comparison. The last column lists the total execution time.<br />

Image Size No. of Iterations Execution time of each Total execution<br />

iteration (seconds) time (hours)<br />

Car 64 × 64 226881 0.80 50.2<br />

Pentagon 32 × 32 83928 0.33 7.6<br />

Overlap 32 × 32 104531 0.40 11.7<br />

Separate 32 × 32 266430 0.39 28.9<br />

Table 6.1: Table of Running Times<br />

In the above simulations, the tolerance parameter for LSQR is chosen to be 10 −10 [117].<br />

The minimum tolerance for Newton’s method is chosen to be 10 −7 . (The Newton iteration<br />

will stop if the norm of the gradient vector is less than 10 −7 .)<br />

6.7 Software<br />

We plan to develop a software package (perhaps named “ImageAtomizer”, following Shaobing<br />

Chen’s “Atomizer” package). At present it includes the following directories:<br />

• Datasets—data files of <strong>image</strong>s and of Matlab functions providing access to them.<br />

• MEXSource—CMEX source code for the direct edgelet transform, the fast approximate<br />

edgelet transform, 2-D DCT, and more.<br />

• Methods—Matlab functions carrying out our global minimum l 1 norm algorithm,<br />

Matching Pursuit, Method of Frames, together with some supporting files.<br />

• Transforms—Matlab functions for the curvelet transform, DCT, edgelet, etc.

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

Saved successfully!

Ooh no, something went wrong!