Texture analysis in sonar images

In the analysis of high-resolution sonar imagery of the seabed, it is important to accurately model the image texture of the sediments in order to detect objects lying on the sea floor or to segment the sediment regions within the image. The received signal is a stochastic process, as it is the summation of a large number of randomly phased echoes that interfere constructively and destructively. This thesis attempts to model the amplitude statistics of sonar imagery with models such as the K-distribution. This has been successfully applied in the past to high-resolution maritime radar clutter and synthetic aperture radar imagery of land scenes. The K-distribution provided a poor fit to images from a number of different sidescan sonar systems. This is probably due to the manufacturer's suboptimal choice of bandwidth of the post-detection filter that resulted in oversmoothing and integration of the sonar returns. Accounting for this leads to the generalised-K distribution that was developed to model weak scattering. This provided an excellent fit to all of the data tested. A small window was passed across the image and the parameters of the generalised-K distribution were estimated from the samples within the window. These parameters were input to a supervised Bayesian classifier that had previously been trained on a number of different sediments. Spatial correlation within the samples was also exploited by Maximum Likelihood spectral classification. The central pixel of the window was then classified as belonging to a particular sediment. A three-class and a four-class test image were both accurately classified. The delineation of a sediment boundary was imperfect due to the finite size of the sampling window, so a method of refining the boundary position using adaptive windows is discussed.