An important problem in high level image analysis is the identification of objects with variable shape. In many applications, the difficulty of this task is compounded by variable lighting effects. In this paper we discuss a specific application to aquaculture, where the objective is to monitor fish populations from images taken with an underwater camera. We investigate the problem of identifying a single fish in a still image. For related work involving stereo image-pairs, see [ 13 ].
The basic approach taken is a Bayesian statistical analysis. The Bayesian paradigm has two key components: the prior and the likelihood. Prior information is available from a set of training data on the size and shape of fish and can be fitted using a Point Distribution Model, as described in § 3 . The likelihood describes how a particular object appears in an image and is much more difficult to model effectively. The effects of variable lighting on the fish images are described in § 2 and an approximate likelihood in terms of edges in the image is described in § 4 .
In mathematical terms, Bayes' Theorem gives the posterior density
for a set of variables
given a set of observations
:
where
is the likelihood of the data and
is the prior density for the variables
which include the location, size, orientation and shape of a fish. The
maximum
a posteriori
(MAP) estimate of
is obtained by maximizing Eq.
1
. The estimation of
by deterministic and Markov chain Monte Carlo (MCMC) techniques is
described in §
5
. The strengths and weaknesses of the MCMC analysis are presented in §
6
and directions for future work discussed in §
7
.
K De Souza
