next up previous index
Next: Simultaneous Segmentation and Up: Unsupervised Estimation with Previous: Unsupervised Estimation with

6.2.1

Simultaneous Restoration and Estimation

 

 

Let us consider a heuristic procedure for simultaneous image restoration and parameter estimation [Besag 1986]. Our interest is to restore the underlying MRF realization, f, from the observed data, d. During this process, two sets of parameters are to be estimated: a set of the MRF parameters and a set of the observation parameters . The procedure is summarized below:

  1. Start with an initial estimate of the true MRF, and guess for and .
  2. Estimate by .
  3. Estimate by .
  4. Update by based on the current , and .
  5. Go to 2 for a number of iterations or until is approximately converged.
This is illustrated below.

Assume that the data is determined by the additive noise model

where is the independent zero-mean Gaussian noise corrupting the true label . Then, for step 2, we can maximize

 

where . Step 2, which is not subject to the Markovianity of f, amounts to finding the ML estimate of which is, according to (6.5),

Steps 3 and 4 have to deal with the Markovianity. Consider again the homogeneous and isotropic auto-logistic   MRF, characterized by (6.7) and (6.8). For step 3, we can maximize

 

where . Heuristics may be used to alleviate difficulties therein. For example, the likelihood function in step 3 may be replaced by the pseudo-likelihood (6.14) or the mean field approximation. A method for computing the maximum pseudo-likelihood estimate of has been given in Section 6.1.2. The maximization in step 4 may be performed by using a local method such as ICM to be introduced in Section 8.2.1.

Fig.6.5 shows an example. The true MRF, f, (shown in upper-left) is generated by using a multi-level logistic   model, a generalization to the bi-level auto-logistic MRF, with a number of M=6 (given) level labels in and the parameters being (known throughout the estimation) and , respectively. The MRF is corrupted by the addition of Gaussian noise with variance , giving the observed data, d, (upper-right). With the correct parameter values and known, no parameters are to be estimated and only step 3 needs to be done, e.g. using ICM. ICM, with fixed, converges after six iterations and the misclassification rate of the restored image is 2.1%. The rate reduces to 1.2% (lower-left) after eight iterations when is gradually increased to over the first six cycles. With known estimated, the procedure consists of all the steps but step 2. The error rate is 1.1% after eight iterations (lower-right), with . When is also to be estimated, the error rate is with and .

  
Figure 6.5: Image restoration with unknown parameters. (Upper-left) True MRF. (Upper-right) Noisy observation. (Lower-left) Restoration with the exact parameters. (Lower-right) Restoration with estimated. From (Besag 1986) with permission; © 1986 Royal Statistical Society.



next up previous index
Next: Simultaneous Segmentation and Up: Unsupervised Estimation with Previous: Unsupervised Estimation with