A Bayesian approach to object detection in sidescan sonar

We consider the problem of object detection against a textured background, and in particular the detection of objects in sidescan sonar. The data set is a series of sidescan segments consisting of an unknown number of irregularly shaped objects in unknown areas of texture. We attempt, through an investigation of the statistical and geometric properties of the data to identify the regions of different textures and the locations of the objects simultaneously, the single point statistics of the various texture classes being known. We consider object detection as a Bayesian image restoration task and propose a model using Gibbs field structures to model the prior knowledge of object placement and a simplified model of image formation. In addition to providing a formal framework for introduction of multiple sources of information, this technique also allows the complexity of modelling to be controlled by specification of the whole model through a set of cooperating submodels. The aim of the technique is to provide a robust object detection system, but also to develop a method for approaching the structuring of complex problems. The use of Monte Carlo Markov chain (MCMC) techniques, geometric structures and relevant parameterisations are proposed as such a method, with advantages in simplicity of model specification and ease of implementation.