Bits per Pixel Requirements for Optimum Encoder
For the processing chain depicted in Fig 1, considering the encoder
operation is preceded and followed by arbitrary linear filter operations
H(f1f2) and P(f4f2) respectively, the following relations, relevant to the
Fourier transformed signal representation, apply:
IN(f1f2) = Fy(f1£2) + Ny(£1£2) (8)
UN(f1f2) 7» HN(£4£2) . IN(£4£2) (9)
CN(f1f2) 7 Un(£9£2) - Dn(£4£2) (10)
ON(E1f2) = By(£1£2) . Oq(f1f2) (11)
where capitals denote the symmetrical two dimensional Fourier Transform of
size N2 and DN(f4f2) denotes the random distortion variable introduced
by the encoder. Applying a minimum mean square quantizer [2], the
following relations are valid (omitting the variables f1f2 for brevity):
E[Dy] = 0 (12)
E[DN-CN.] = 0 (13)
|]Cy//2 = max [0, 0/1 2 — (Dy((2] (14)
2 1
HN OA
E [Ny.Dy.] = RTE (15)
_ Ent Ball
E[FNDN] = Hz (16)
The distortion measure dp of the processed image € (x,y) with respect to
the reference image f(x,y) is defined here by the squared error criterium.
Hence:
dr = E[£ € (xy) - f(x,y)) ?] (17)
or equivalenty,
dp- 1 C(t £9) = F(f1f 2 18
T weld 1£2) - F(£1£2)l1 (18)
Applying a minimum mean square encoder for each of the Fourier
coefficients it follows after some manipulation (omitting var. f 1£2):
^J 2 2. à 2 t Fu -— = +
IIc El TRITT [P) [H[ I| ll eu (Pu ePi )-Iib [P]
IU, à à
~ LN . 1 (iJ ll (| D I!
JIC-FI} = Il FIL ce lor < Uo wl SUD,
The upperbound of the minimum average number of bits per pixel R(dg) as
function of the encoder induced distortion dg is given by the rate
distortion function of a corresponding Gaussian process and is expressed
parametrically by [3]:
2,
R(de) = a [max [0,104 ze Ja (20)
Sa iT
de . Jf "^ [eur 5 In ace [ 4. a f, (21)
448
La ax a RÉ A Se JP PER S de
PA