Spatial Interaction Models For Sonar Image Data

This paper describes an approach to sidescan SONAR data based on texture analysis of images, where the segmentation is constructed using stochastic descriptions of the data derived directly from the spatial domain. The paper outlines the spatial analysis technique, and describes a number of models that may be used. The stages in model construction are considered with attention to estimates of reliability and accuracy in parameter estimation, and model selection. Finally, a segmentation technique based on generating a realisation of a random process is described, and a number of examples are considered. 2. SPATIAL MODELS 2.1. Conditional Probability models of texture Texture analysis in general has a voluminous literature. Many statistical techniques have been used, along with fractals, power spectra and other ideas. Recently, analysis of texture in terms of a probability structure measured on the pixel values has become more common [14][13][10], and provides a formal model of texture with a very rich mathematical structure. The authors have previously defined a segmentation model based on splitting an image into a number of binary planes and analysing each one separately [18]. The model here retains this idea but extends the models of the binary processes on each plane by describing the local conditional probability distribution structure explicitly, that is, the probability of a pixel being a particular class (or greylevel, etc.) given a finite set of neighbours. Conditional probability as a model for spatial data was proposed by Besag [1] for analysis of lattice systems. The conditional distribution was more appealing that the equivalent joint distribution, he argued, partly because of its intuitively simple application to typical problems ("what is the probability th..