Time frequency and spatial feature generaton for food kernel inspection and brain computer interface

We propose feature generation methods that provide better generalizations in the two application domains of interest. The particular applications we are interested in are food kernel inspection via impact acoustics and Brain Computer Interface (BCI). Both of the problems involve binary classifications. In the first problem the goal is to separate closed pistachio nuts from those of open ones with only a few input features using the acoustic signals as the input. The second problem involves a BCI application where the goal is to achieve accurate classification of a motor task between two classes using the multi channel brain signals as the input. In order to achieve the goal in the first problem, we propose to generate a rich time frequency feature dictionary from the impact acoustics signals using a double tree wavelet transform. In the next step, a few most important features are selected via a search method. Experimental results show that the performance of the proposed approach is superior to that of traditional approaches. Furthermore, with this approach the dependence on prior information is minimal. We propose two separate feature generation approaches for the second problem. The first approach exploits the time and frequency content of the multi channel neural recordings by employing the same double tree approach used in first problem. A structured redundant feature dictionary is generated for each spatial location separately. In order to achieve better generalization a small subset of features are selected from the large feature dictionary. Experimental results using the data set of BCI competition 2005 show that a better performance is achievable with a small number of features without manual selection of sensors. In the second approach, we consider the feature generation method known as Common Spatial Pattern (CSP) that exploits the spatial dimension of the multi channel input. We propose to implement a sparse CSP method for feature generation using a greedy search based generalized eigenvalue decomposition. We show the performance gain obtained by sparse CSP based approach over its traditional solution on the 2005 brain computer interface competition datasets. The results show that improved classification performances are achievable with reduced complexity.