20.01.2013 Views

motion estimation and compensation for very low bitrate video coding

motion estimation and compensation for very low bitrate video coding

motion estimation and compensation for very low bitrate video coding

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

184 Chapter C. Markov R<strong>and</strong>om Fields<br />

the estimate of the solution u according to the collected data d:<br />

where p(d)= P u p(dju):p(u).<br />

p(ujd)= p(dju):p(u)<br />

p(d)<br />

(C.1)<br />

Bayesian modeling is often used to determine the Maximum A Posteriori<br />

(MAP) estimate, i.e. the value of the solution u that maximizes<br />

the conditional probability p(ujd). The Maximum Likelihood (ML)<br />

estimate does not statistically describe u: it considers it as a vector of<br />

parameters. The ML is a special case of the MAP, in which the a priori<br />

model p(u) is constant (uni<strong>for</strong>m distribution). The aim then becomes<br />

to maximize the conditional probability p(dju).<br />

Working with the logarithm of the posterior density, one obtains:<br />

log(p(ujd)) = log(p(dju)) + log(p(u)) , log(p(d)); (C.2)<br />

where the last term does not depend on u <strong>and</strong> can be neglected in<br />

maximization processes with respect to u. The MAP estimate is thus:<br />

[ @log(p(dju))<br />

@u<br />

+ @log(p(u))<br />

]ju=^uMAP =0; (C.3)<br />

@u<br />

where u =^uMAP is the solution at the maximum. Similarly, the ML<br />

estimate is, with u =^uML the solution at the maximum:<br />

C.2 Markov R<strong>and</strong>om Fields<br />

[ @log(p(dju))<br />

]ju=^uML =0: (C.4)<br />

@u<br />

Let consider an image I, in which pixels de ne a lattice S of sites s:<br />

S = fs =(i; j)g (C.5)<br />

To e<strong>very</strong> site s is associated a r<strong>and</strong>om variable [99] As, whose values<br />

as belong to . For instance, = f0; :::; 255g represents the possible<br />

luminance values of a black <strong>and</strong> white picture <strong>and</strong> = f0; :::; 255g q the<br />

values that can be associated with any pixel of a mutispectral image<br />

with q channels.

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

Saved successfully!

Ooh no, something went wrong!