A novel feature selection approach applied to underwater object classification

A novel filter method for feature selection is presented. In our research, we observed that the feature relevance measures in the literature evaluate the features for classification purposes only with respect to certain aspects, e.g. distance, information theory, etc. Accordingly, the resulting feature selections may only be adapted to a narrow range of classifiers. Our approach jointly considers two relevance measures, i.e. mutual information (MI) and Relief weight (RW) so that the features are assessed more comprehensively. It requires not only the selection to hold sufficient MI, it also forces the features in the selection to have large RWs. In order to avoid an NP hard problem, a heuristic searching scheme is adopted, i.e. sequential forward searching. Moreover, the selection's cardinality can be determined automatically. Finally, this approach is applied to the underwater object classification and its classification results are compared to those of filter methods in the literature.