[Cover] [Contents] [Index] |
Page 248
Once the prior energy function formulation is established, the second step is to compute the likelihood energy. For an image contaminated by independent identical distribution (i.i.d) Gaussian noise er with zero mean and variance σ2, the observation dr can be expressed as dr=wr+er, where wr is the true pixel intensity (grey) value, and er is the i.i.d. Gaussian noise. The probability P(dr|wr) and likelihood energy U(dr|wr) can be expressed as:
(6.27) |
and the posterior energy U(w|d) can be derived as:
(6.28) |
or, alternatively, as:
(6.29) |
where q is a weighting parameter such that the larger the value of q, the more the solution will be affected by the second term, i.e. the prior energy.
If the smoothness prior energy U(w) takes a quadratic form as shown above, U(w) is also called a regulariser. Without this regulariser, one or more of the following conditions may apply (Tikhonov and Arsenin, 1977): (1) a solution does not exist; (2) the computed solution is not unique; or (3) the computed solution does not depend on the initial state. In such circumstances, the problem is ill-posed. However, the use of this regulariser can convert an ill-posed problem into well-posed problem, i.e. guarantee that a unique and initial-state-dependent solution is obtained.
The simplest way to minimise the energy function in Equation (6.29) is to use a gradient descent minimisation algorithm (some alternative minimisation algorithms are discussed in Section 6.4). Given an initial configuration w(0), then, at iteration i, the new value for wr(i+1) at site r is obtained from:
(6.30) |
or equivalently (by using Equation (6.24)):
[Cover] [Contents] [Index] |