Subset selection from biased dictionaries for impact acoustic classification

In this paper we study a sparse signal representation approach for the classification of impact acoustic signals obtained from empty and full hazelnuts. In particular, two custom dictionaries are designed for each class with a vector quantization algorithm by using the training data for each. In the following step each individual dictionary or their combination is used for representing the test acoustic signals with the belief that the representation will be biased. Two different subset selection (SS) techniques, matching pursuit (MP) and a bounded error subset selection algorithm (BESS) were investigated to approximate given signals by using the code vectors in these biased dictionaries. The approximation error and the number of code vectors selected from dictionaries were used as input features for the final classification step. We observe that this biased dictionary design allows one to distinguish between classes while representing them with a small number of codevectors. In particular the classification performance of the BESS algorithm by using the codevectors constructed from empty hazelnut acoustics was the best and outperforms that of using the MP algorithm. The combination of BESS algorithm, the dictionary derived from empty hazelnut acoustics and a decision tree yields classification accuracies of 94.8% and 100% for empty and full hazelnuts respectively. Our results indicate that subset selection from biased dictionaries can be used as a new approach for classification.