Mapping interesting regions in qualitative sidescan sonar imagery predominantly relies on an expensive human interpretation process. It would therefore be useful to automate components of this task with a feature-based, Machine Learning system. We must first establish a framework for reliably and efficiently evaluating the features. A novel ensemble of probabilistic distance measures is proposed, as an objective function for this purpose. The idea is motivated by the fact that different distance measures yield conflicting feature ranking results. In the ensemble, distances can be combined to produce a consensus rank score. As a test case, we find a sub-optimal parameterisation of a Co-occurrence Matrix, for identifying textures peculiar to the tube-building worm, Sabellaria spinulosa. A strong correlation is found between ensemble scores and classification accuracies. The proposed methodology is applicable to any sonar imagery, classification task or feature groups.
|Title of host publication||Intelligent Data Engineering and Automated Learning - IDEAL 2011|
|Editors||Hujun Yin, Wenjia Wang, Victor Rayward-Smith|
|Number of pages||10|
|Publication status||Published - 2011|