Scribd
Upload
197 views

Wiener Helstrom Filter

The document discusses the Wiener-Helstrom filter, which is used for image restoration. It derives the optimal filter by minimizing the mean square error between the estimated and actual input images. The filter is derived using variational calculus and by considering small perturbations. The solution shows that the optimal filter is the ratio of the cross-power spectrum of the ideal and blurred images to the total power spectrum. This filter minimizes the mean square error between the estimated and actual input images.

Uploaded by

Mina Mikhael
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
197 views

Wiener Helstrom Filter

The document discusses the Wiener-Helstrom filter, which is used for image restoration. It derives the optimal filter by minimizing the mean square error between the estimated and actual input images. The filter is derived using variational calculus and by considering small perturbations. The solution shows that the optimal filter is the ratio of the cross-power spectrum of the ideal and blurred images to the total power spectrum. This filter minimizes the mean square error between the estimated and actual input images.

Uploaded by

Mina Mikhael
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

Fundamentals of Digital Image Processing a practical approach with examples in Matlab Chris Solomon and Toby Breckon

The Wiener-Helstrom filter The (linear, shift invariant) imaging equation is expressed in the spatial domain as

g x, y

f x , y h x x , y y dx dy

n x, y

(1)

where g x, y is the recorded image, f x, y is the corresponding input distribution,


h x x,y y

is the point spread function and n x, y is the additive noise.

Use of the convolution theorem gives the frequency domain equivalent directly as
G kx , k y H kx , k y F kx , k y N kx , k y

(2)

The basic image restoration problem is to form some estimate of the input distribution f x, y that approximates the actual distribution f x, y as closely as possible. An intuitive and well-defined measure is to choose f x, y so as to minimise the overall mean square error defined as
Q f x, y
2

f x, y

dxdy

(3)

where the angle brackets denote averaging over any stochastic variation in the estimate f x, y . The Wiener-Helstrom filter starts by transforming equation (3) into its frequency domain equivalent. Parsevals theorem states that
2 2

f x, y

dxdy

F kx , k y

dk x dk y

(4)

Applying this to the two terms in eq. (3), we therefore obtain


Q F kx , k y
2

F kx , k y

dk x dk y

(5)

The Wiener-Helstrom filter is linear and we therefore posit a filter Y k x , k y such that our estimate is formed by multiplying the output spectrum by the filter F kx , k y Y kx , k y G kx , k y

(6)

Fundamentals of Digital Image Processing a practical approach with examples in Matlab Chris Solomon and Toby Breckon

Substituting equations (2) and (6) into equation (5) (and dropping the explicit dependence of each term on kx and k y for brevity), we obtain
Q dk x dk y HF N Y *Y HF
*

Y HF

N F * F HF

N Y * F *F

(7) We note at this juncture that the input distribution and the noise are the unknown * quantities and thus treated as stochastic. Making use of the relation AB B* A* equation (7) may be expressed as
Q Q dk x dk y dk x dk y F *H * N * Y *Y HF N Y HF N F * F F *H * N *Y *YN F *F N * Y * F *F

F * H *Y *YHF YHFF *

F * H *Y *YN FF * H *Y *

N *Y *YHF FN * Y *

Y NF *

(8) We can factor out the non-stochastic terms from the averaging brackets to write this as
Y *Y H Q dk x dk y
2

F *F

Y *YH * F * N

Y *YH N * F Y * FN *

Y *Y N * N F *F

YH FF *

Y NF *

H *Y * FF *

(9)

It is a reasonable assumption to assume that there is no statistical correlation between the unknown input distribution represented by F and the noise process represented by N. Certainly, there is no physical basis for this in real imaging systems. Accordingly, we may set NF * FN * 0 . Equation (8) then reduces to
Q dk x dk y Y *Y H WF Y *YWN YHWF
2

H *Y *WF WF
2

(10)

Where the noise power spectrum WN k x , k y


WF k x , k y F kx , k y
2

N kx , k y

and the input power spectrum

There are two possible (and in fact equivalent) routes to minimising Q in equation 9 for the unknown filter Y. For the reader familiar with methods of variational calculus, we note that the integrand in eq (9) is a functional, in which we seek that function Y which will minimise the integral. In general this functional may have explicit dependencies on Y, derivatives of Y and the spatial frequency variables kx , k y , we thus write -

Fundamentals of Digital Image Processing a practical approach with examples in Matlab Chris Solomon and Toby Breckon

L Y k x , k y , Yk x , Yk y ; k x , k y where Ykx Y Y and Yk y kx ky

Y *Y H WF Y *YWN YHWF

H *Y *WF WF

(11)

We can then use the standard Euler-Lagrange solution for a single function of 2 variables, namely
L Y* d dk x L Yk x d dk y L Y 0

(12)

Since the functional L has no explicit dependence on Ykx and Yk y , this reduces simply to

L Y
Thus,

(13)

L Y *Y H WF L Y Y * H WF
2

Y *YWN YHWF Y *WN HWF 0

H *Y *WF WF

(14)

Yielding the solution

H *WF H WF WN
2

(15)

A second more direct approach to the solution in this particular case is to consider small perturbation/variations in the function Y, Y Y Y and demand that the corresponding variation in the functional as L L L is stationary (i.e. zero). Starting from equation (11), we thus have L Y *Y H WF Y *YWN YHWF L Y * Y H WF YWN
2 2

H *Y *WF WF Y Y * H WF Y *WN
2

H *WF

(16)
HWF 0

which yields an identical result directly

H *WF H WF WN
2

(17)

Fundamentals of Digital Image Processing a practical approach with examples in Matlab Chris Solomon and Toby Breckon

Substitution of eq. (17) back into eq. (10) shows that the mean-square error obtained is
H WF2 H WF WN
2 2

dk x dk y

H WF2 H WF WN
2

H WF2 H WF WN
2

WF

H WF WN H WF WN
2

which simplifies to
WFWN H WF WN
2

dk x dk y

(18)

You might also like