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

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.1. DCT AND HOMOGENEOUS COMPONENTS 39<br />

Type of DCT δ 1 δ 2 transform matrix<br />

type-I DCT 0 0 CN<br />

I<br />

type-II DCT 0 1/2 CN<br />

II<br />

type-III DCT 1/2 0 CN<br />

III<br />

type-IV DCT 1/2 1/2 CN<br />

IV<br />

Table 3.2: Definitions of four types of DCTs.<br />

For all four types, the DCT matrices are orthogonal (C −1<br />

N<br />

of course they are unitary. Another noteworthy fact is that C II<br />

N<br />

= CT N<br />

). Since they are real,<br />

=(CIII N )−1 .<br />

We can view the four types of DCT as choosing the first index (which usually corresponds<br />

to the frequency variable) and the second index (which usually corresponds to the time<br />

variable) at different places (which can be either integer points or half integer points).<br />

Figure 3.1 depicts this idea.<br />

DCT <strong>via</strong> DFT<br />

Particularly for the type-I DCT, the transform matrix CN I is connected to the transform<br />

matrix of a 2N-point DFT (DFT 2N ) in the following way:<br />

⎡⎛<br />

[ ]<br />

C<br />

I<br />

N<br />

•<br />

= R<br />

• • ⎢⎜<br />

⎣⎝<br />

1 √<br />

2<br />

1<br />

. ..<br />

1<br />

⎞ ⎛<br />

DFT<br />

⎟ 2N ⎜<br />

⎠ ⎝<br />

1 √<br />

2<br />

1<br />

. ..<br />

1<br />

⎞⎤<br />

,<br />

⎟⎥<br />

⎠⎦<br />

where R is a real operator: R(A) is the real part of matrix A. Symbol “•” represents an<br />

arbitrary N × N matrix.<br />

The other types of DCT, generally speaking, can be implemented <strong>via</strong> the 8N-point DFT.<br />

We explain the idea for the case of type-IV DCT. For type-II and type-III DCT, a similar<br />

idea should work. Note that in practice, there are better computational schemes.<br />

Let us first give some mathematical notation. Let symbol I N→8N denote an insert<br />

operator: it takes an N-D vector and expand it into an 8N-D vector, so that the subvector<br />

at locations {2, 4, 6,... ,2N} of the 8N-D vector is exactly the N-D vector, and all other

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

Saved successfully!

Ooh no, something went wrong!