Underwater target detection in synthetic aperture sonar data

The detection of underwater targets, such as mines, from sonar returns is a difficult task which is compounded by the complex and variable backgrounds found on the seabed. The developed system employs a classical training and classification structure giving a statistical characterisation of the background together with domain knowledge of typical target types. A set of ground truth labels have been produced for three given seabed test regions which contain a range of target types. The method identifies the centre of targets using log-Gabor, matched and shaped filters together with a Support Vector Machine (SVM) classifier. Subjective testing enabled the comparison of our automatic detection methods with the performance of expert operators. The automatic target detection method was found to offer performance at least as good as human operators on identical data (based on a small operator data set).