Bayesian approach to object detection in sidescan sonar

The authors consider the problem of detecting discrete objects in sidescan sonar, utilizing the geometric and statistical properties of the objects expressed in a Bayesian framework. The model proposed hypothesizes that the observed image is formed through observations of a simple texturing process, modified to incorporate the presence of potential objects. Suitable priors are selected expressing geometric and spatial constraints, and a Markov chain Monte Carlo (MCMC) system is used to estimate texture labels and probability of object presence on a per pixel basis. The model and the method of parameter construction are described. Some examples of typical object detection are given, along with the results of a more detailed study on a groundtruth data set. It is concluded that the model proposed appears effective for this data set, and is flexible enough to be easily trained for use in others. Groundtruth object detection with a correction detection of 87% is observed, with a false alarm rate of 0.19/image.