Many image analysis and computer vision problems can be formulated as a scene labeling problem. Bayesian modeling of images by Markov random fields is a coherent theoretical framework. It has however some drawbacks, one of which is the computational complexity. Because the energy function has many local minima, most deterministic or local optimization algorithms depend on the starting point, i.e., the better the initialization, the bigger the chance of the final result close to the global optimum. Usually, the initialiation uses maximum likelihood estimation (MLE) for each site and it is not good enough in practice. We propose two approaches to obtain better initialization than the traditional MLE, one is based on circular window sampling, another is "spotlight" operator. From the experiments, we can see the two approaches are very effective and efficient for initializations, and the fast ICM optimization based on them can provide satisfactory labeling results.
[1]
William E. Higgins,et al.
A flexible implementation of maximum-homogeneity filtering for 2-D and 3-D images
,
1991,
IEEE Trans. Signal Process..
[2]
J. Besag.
On the Statistical Analysis of Dirty Pictures
,
1986
.
[3]
Alan L. Yuille,et al.
Region Competition: Unifying Snakes, Region Growing, and Bayes/MDL for Multiband Image Segmentation
,
1996,
IEEE Trans. Pattern Anal. Mach. Intell..
[4]
Stan Z. Li,et al.
Markov Random Field Modeling in Computer Vision
,
1995,
Computer Science Workbench.
[5]
J. Besag.
Spatial Interaction and the Statistical Analysis of Lattice Systems
,
1974
.
[6]
Donald Geman,et al.
Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images
,
1984
.