Subset selection with structured dictionaries in classification

This paper describes a new approach for the selection of discriminant time-frequency features for classification. Unlike previous approaches that use the individual discrimination power of expansion coefficients, the proposed approach selects a subset of features by implementing a classifier directed pruning of an initial redundant set of candidate features. The candidate features are calculated from a structured redundant time-frequency analysis of the signal, such as an undecimated wavelet transform. We show that the proposed approach has a performance that is as good as or better than traditional classification approaches while using a much smaller number of features. In particular, we provide experimental results to demonstrate the superior performance of the algorithm in the area of impact acoustic classification for food kernel inspection. The proposed algorithm achieved 91.8% and 98.5% classification accuracies in separating open shell from closed shell pistachio nuts and discriminating between empty and full hazelnuts respectively. Traditional methods used in this area resulted in 82% and 97% classification accuracies respectively.