Bayesian Image and Video Enhancement Using a Non-Gaussian Prior - Nonlinear Signal and Image Processing: Theory, Methods and Applications - page 313

(a) (b)

(c) (d)

FIGURE 9.8
Region-of-interest image magnification of the IKONOS satellite picture with an effective GSD of
4 m. (a) Microview of the original 1-m image data. (b) 4-m image data. (c) Bilinear interpolation
of the 4-m pixels. (d) Cubic B-spline interpolation of the 4-m pixels. (e) Bayesian enhancement of
the 4-m data using a Gauss-Markov random field model. (f) Bayesian enhancement of the 4-m
data using a Huber-Markov random field model with T H = 5. Note that the nonlinear estimate

generated using the HMRF model is slightly sharper than the linear interpolation generated by
the GMRF model. ( Continued )

1 .

with a different set of constraints required to satisfy the block-DCT coding
observation model:

X = x : y

H − 1 Q [ Hx ] .

=

(9.42)

The original, observed image data x 0 y is used as the initial condition for

the gradient projection algorithm. Because the compressed data is in the block

=