Computational Vision and Medical Image Processing IV

92

(Haar)). The observation matrix R, of dimen-

sion k × k, is a Toeplitz blocks matrix represent-

ing 2D convolutions. The observed image (figure

(3)) is obtained convolving the original image

(figure (3)) with a uniform blur filter of size 9 × 9,

and then adding to the blurred image (Rx) white

Gaussian noise with σ = 0.005. Taken as initial

estimate

(

0

)

T

ˆ

φ

=

y

x

, figures (3) - (3) shows the

algorithm performance for q = 2 and p ∈ ]0, 1]. For

q = 2 ^ p = 0.5 the initial value of the objective func-

tion is 2.0834 × 10

6

and the final one is 3.146 × 10

4

.

A typical statistical model for image wavelet coeffi-

cients is the Generalized Gaussian Density [21]. In

this experiment the best approximation occurs for

p = 0.5. This is consistent with Moulin ...