Skip to Main Content
One of the main challenges in mine detection using sidescan sonar images is the high density of mine-like objects (MLOs) in a clutter environment. This paper proposes an image change detection technique for bitemporal images which suppresses false alarms efficiently without involving large training data sets. The proposed approach uses the spatial dependence of a stationary object between bitemporal images to eliminate the differences caused by a position error. Bayes theory is then employed to classify the changed and unchanged objects. In particular, the a priori probabilities are formulated by the Markov random field (MRF). The likelihood functions are modeled using the coarseness difference of objects as the test statistics, and the parameters are estimated using the expectation-maximization (EM) method. Real sidescan sonar data are used to validate the proposed method. Results show that the proposed MRF change detection method is robust to the poor quality of object boundaries due to speckle noise, and outperforms the conventional pixel-level change detection methods.