Segmentation of seafloor sidescan imagery using Markov random fields and neural networks

Segmentation of seafloor acoustic imagery is of great importance in a wide range of applications. Although considerable successes have been achieved, a critical issue in this domain is the lack of reliable 2D image models from which segmentation and related processing can consistently proceed. The authors describe a nonparametric algorithm for the segmentation of seafloor sidescan imagery (SSI) based on a combination of Markov random fields and multiple layer perceptrons (MLP). SSI, which is considerably noisy and textured, is embodied by a hierarchy of Gibbs distributions. Segmentation of SSI is then considered as a maximum a posteriori estimation. To obtain better estimates of local likelihoods, an MLP is adopted for learning the distribution of observations from training data. Experimental results using data from a Pacific midocean ridge area are demonstrated