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.

84 CHAPTER 4. COMBINED IMAGE REPRESENTATION<br />

with some recent advances in theoretical study, show that a global optimization algorithm<br />

like BP is more stable in recovering the original <strong>sparse</strong> decomposition, if it exists. But BP<br />

is a computationally intensive method. The remainder of this thesis is mainly devoted to<br />

overcoming this barrier. In the next section, Section 4.3, we first explain the optimality of<br />

the minimum l 1 norm decomposition and then give our formulation, which is a variation<br />

of the exact minimum l 1 norm decomposition. This formulation determines the numerical<br />

problem that we try to solve.<br />

4.3 Minimum l 1 Norm Solution<br />

The minimum l 1 norm solution means that in a decomposition y =Φx, where y is the desired<br />

signal/<strong>image</strong>, Φ is a flat matrix with each column being an atom from an overcomplete<br />

dictionary and x is a coefficient vector, we pick the one that has the minimum l 1 norm<br />

(‖x‖ 1 ) of the coefficient vector x.<br />

A heuristic argument about the optimality of the minimum l 1 norm decomposition is<br />

that it is the “best” convexification of the minimum l 0 norm problem. Why? Suppose we<br />

consider all the convex functions that are supported in [−1, 1] and upper bounded by the<br />

l 0 norm function and we solve<br />

⎧<br />

⎪⎨ f(x) ≤‖x‖ 0 ,<br />

maximize f(x), subject to ‖x‖ ∞ ≤ 1,<br />

x<br />

⎪⎩<br />

f is convex.<br />

The solution of the above problem is ‖x‖ 1 .<br />

The ideas of using the minimum l 1 norm solutions in signal estimation and recovery date<br />

back to 1965, in which Logan [96] described some so-called minimum l 1 norm phenomena.<br />

Another good reference on this topic is [55]. The idea of the minimum l 1 norm phenomenon<br />

is the following: a signal cannot be “<strong>sparse</strong>” in both time and Fourier (frequency) domain;<br />

if we know the signal is limited in one domain (e.g., frequency domain) and the signal is<br />

unknown in a relatively small set in another domain (e.g., time domain), then we may be<br />

able to perfectly recover the signal by solving the minimum l 1 norm problem.<br />

This phenomenon is connected to the uncertainty principle. But the conventional uncertainty<br />

principle is based on the l 0 norm. In the continuous case, the l 0 norm is the length<br />

of interval; in the discrete case, the l 0 norm is the cardinality of a finite set. As we know,

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

Saved successfully!

Ooh no, something went wrong!