Markov Random Field Models in Computer Vision: A Posteriori

Markov R a n d o m Field Models
in C o m p u t e r Vision
S.Z. Li
School of Electrical and Electronic Engineering

Nanyang Technological University, Singapore 2263
[email protected]
A b s t r a c t . A variety of computer vision problems can be optimally

posed as Bayesian labeling in which the solution of a problem is de-
fined as the maximum a posteriori (MAP) probability estimate of the
true labeling. The posterior probability is usually derived from a prior
model and a likelihood model. The latter relates to how data is observed
and is problem domain dependent. The former depends on how various
prior constraints are expressed. Markov Random Field Models (MRF)
theory is a tool to encode contextual constraints into the prior proba-
bility. This paper presents a unified approach for MRF modeling in low
and high level computer vision. The unification is made possible due to a
recent advance in MRF modeling for high level object recognition. Such
unification provides a systematic approach for vision modeling based on
sound mathematical principles.
1 Introduction
Since its beginning in early 1960's, computer vision research has been evolving
from heuristic design of algorithms to systematic investigation of approaches
for solving vision problems. In their search for solutions, researchers have real-
ized the importance of contextual information in image understanding. In this
process, a variety of vision models using context have been proposed. Among
these are Markov Random Field (MRF) theory based models (of which analytic
regularization theory based models are special cases).
MRF modeling is appealing for the following reasons (Preface of [4]): (1)
One can systematically develop algorithms based on sound principles rather
than on some ad hoc heuristics for a variety of problems; (2) It makes it easier
to derive quantitative performance measures for characterizing how well the
image analysis algorithms work; (3) MRF models can be used to incorporate
various prior contextual information or constraints in a quantitative way; and
(4) The MRF-based algorithms tend to be local, and tend themselves to parallel
hardware implementation in a natural way.
Complete stochastic vision models based on MRF are formulated within the
Bayesian framework. The optimal solution of a problem is defined as the max-
imum a posteriori (MAP) probability estimate of the truth, the best that one
can get from random observations. Most of vision problems can be posed as one
of labeling using constraints due to prior knowledge and observations. In this
Lecture Notes in Computer Science, Vol. 801

Jan-Olof Eklundh (Ed,)
Computer Vision - ECCV <)4
9 Springer-Verlag Berlin Heidelberg 1994
362
case, the optimal solution is defined as the MAP labeling and is computed by
minimizing a posterior energy. The posterior probability is derived, using the
Bayesian rule, from a prior model and a likelihood model. The latter relates to
how data is observed and is problem domain dependent. The former depends on
how various prior constraints are expressed. Results from MRF theory provide
us tools to encode contextual constraints into the prior probability. This is the
main reason for MRF vision modeling.
MRF based approaches have been successful in modeling low level vision
problems such as image restoration, segmentation, surface reconstruction, tex-
ture analysis, optical flow, shape from X, visual integration and edge detection
(There are a long list of references. Readers may refer to collections of papers in
[15, 4] and references therein). Relationships between low level MRF models are
discussed in [16, 7] and those between MRF models and regularization models in
[16]. The unifying theme of Bayesian modeling for low level problems appear for
example, in [7, 2, 18]. A prototypical Bayesian formulation using MRF is that
of Geman and Geman [8] for image restoration.
Investigation of MRF modeling in high level vision such as object matching
and recognition, which is more challenging (Introduction of [15]), begins only
recently. In an initial development of an MRF model for image interpretation
[17], the optimal solution is defined as the MAP labeling. Unfortunately, the
posterior probability therein is derived using heuristic rules instead of the laws
of probability, which dissolves the original promises of MRF vision modeling. A
coupled MRF network for simultaneous object recognition and segmentation is
described in [5].
In a recent work [11], an MRF model for high level object matching and
recognition is formulated based ou sound mathematical principles. Mathemati-
cally, like the typical low level MRF model of Gemau and Geman [8], the model
utilizes MRF theory to characterize prior contextual constraints. This, plus an
observation model for the joint likelihood, enables the derivation of the posterior
probability. The model [11] is more general than the low level model [8] in that
it makes use of contextual observations and allows non-homogeneous sites and
non-isotropic neighborhood systems.
This makes it possible to formulate a larger number of low and high level
problems in the single Bayesian framework in a systematic way. This is of signif-
icance in both theory and practice. It provides a rational approach on a sound
basis. It implies some intrinsic properties or common mechanisms in seemingly
different vision problems. It also suggests that these problems could be solved
using a similar architecture.
This paper presents such a unified MRF modeling approach [10]. The sys-
tematic way to the MP~F modeling is summarized as five steps:
1. Pose the vision problem as one of labeling in which a label configuration

represents a solution (See.2).
2. Further pose it as a Bayesian labeling problem in which the optimal solution
is defined as the MAP label configurations (See.3),
3. Use Gibbs distribution to characterize the prior distribution of label config-
urations (Sec.3.2),

Markov Random Field Models in Computer Vision: A Posteriori

Uploaded by

Copyright:

Available Formats

Markov Random Field Models in Computer Vision: A Posteriori

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Markov Random Field Models in Computer Vision: A Posteriori

Uploaded by

Copyright:

Available Formats

Markov R a n d o m Field Models

School of Electrical and Electronic Engineering

A b s t r a c t . A variety of computer vision problems can be optimally

Lecture Notes in Computer Science, Vol. 801

1. Pose the vision problem as one of labeling in which a label configuration

You might also like