Fei T.,Institute of Water Acoustics |
Kraus D.,Institute of Water Acoustics |
Berkel P.,Institute of Water Acoustics
Proceedings of Meetings on Acoustics | Year: 2012
Due to the high quality images provided by modern synthetic aperture sonar systems, automatic target recognition attracts increasing attention. In order to describe the mine-like-objects in sonar images, numerous features are available in the literature to support classification process. This paper introduces a new idea for the feature selection. Since an individual relevance measure considers the classification information only in a certain aspect, the feature relevance should rely on a combination of several relevance measures instead of any single measure. Both linear and nonlinear combinations of the three chosen measures are studied. New feature relevance measures are accordingly proposed and compared. A sequential forward searching scheme is used for maximizing the proposed feature relevance measures. The induced approach is called maximum-composite-relevance-measure-using-sequential-forward-searching-scheme (mCRM-SFSS). Compared to those approaches in the literature, like MIFS, MIFS-U and mRMR, mCRM-SFSS is able to determine the cardinality of the selection automatically without prior knowledge of the length of the feature vector. Furthermore, it substitutes the empirical principle of Occam's razor in RELFSS with the maximum entropy principle, which is proven to be better suited to our application. Finally, mCRM-SFSS is applied to underwater object feature extraction and compared to those of existing approaches. © 2013 Acoustical Society of America.