By Sabrina Matteucci
Probabilistic modelling of sidescan sonar texture using pairwise pixel interactions. Myers, Vincent L. SR-424. June 2005.
Probability models used to describe sidescan sonar background pixels are usually represented using Rayleigh or K-distributions. While useful for many purposes, these distributions do not capture information which characterises non-independent pixel amplitudes which typify many textures found in modern high-resolution sidescan sonar data. Such models are necessary if we are to compute sensor and mission performance metrics such as probabilities of detection and classification. This report describes a flexible technique for describing sidescan sonar image textures using Markov Random Fields. Pairwise pixel interactions are expressed using gray level difference histograms and, due to the equivalence of Markov and Gibbs random fields, are used as sufficient statistics to a Gibbs probability distribution. The parameters of the Gibbs distribution are found using stochastic approximation and complex textures such as sand ripples and sea grass are successfully represented using this method. The Gibbs model is also shown to be capable of seabed segmentation based on texture, as well as capable of designing a simple constant-false-alarm-rate detector.