Noise robust feature extraction for underwater targets classification via label consistent dictionary learning

Underwater targets classification using traditional features suffers significant performance loss under the mismatched noisy condition when the training data and the testing data are measured at totally different noise levels. We consider to denoise the measured signals by the sparse representation. However, the real acoustic signals for underwater targets are too complicated to be represented by a fixed simple dictionary, we suggest to learn the representative dictionary from the data. To simultaneously learn the dictionary and discriminative sparse features, the underlying structure that sparse patterns are similar for sample data from the same target and differ a lot among different targets is exploited by using the label consistent dictionary learning method. The resulted sparse coefficients are then used as the input features of the Support Vector Machine (SVM) for target classification. Experimental results using real underwater acoustic dataset demonstrate that the proposed noise robust feature outperforms the traditional ones under the mismatched noisy condition.