Estimation of Objects In Highly Variable Images Using Markov Chain Monte Carlo

K.M.A. de Souza J.T. Kent K.V. Mardia
Department of Statistics, University of Leeds, LS2 9JT
kds@amsta.leeds.ac.uk

R.D. Tillett
Bioengineeering Division, Silsoe Research Institute,
Wrest Park, Bedford, MK45 4HS

Abstract:

A general approach is presented here for the estimation of objects in digital images using statistical techniques. Unlike various methods in machine vision, the likelihood and prior distribution are explicitly specified. Furthermore the posterior distribution of template variables is examined using Markov chain Monte Carlo analysis and compared with deterministic methods. A point distribution model for the objects is fitted by Procrustes analysis on a set of training images and used to construct a prior distribution for a deformable template. The edges of objects are identified from a smoothed gradient image and the spatial distribution of these edges is used to formulate the likelihood for a proposed template. The strengths and weaknesses of this approach are discussed and results are presented for an application in aquaculture involving underwater images of fish. A key challenge in this application arises from the variable nature of edges due to the variable surface orientation within a fish with respect to the light source. Future developments are outlined.

Next: 1 Introduction

K De Souza
Fri Jul 11 14:17:05 BST 1997