A Multilayer Neural Network for Classification of Frequency Information Dominant Patterns

Features of pattern data can be expressed in a more informative feature domain in order to improve the classification performance. This paper proposes a new implementation of the multi-layer feed-forward neural network that can do classification based on the frequency features extracted by the first hidden layer that performs correlational filter operation. The correlational feature extraction layer performs the filtering operation with the Fourier transformed input pattern, which is resulted in complex data form. The correlation filter output is then converted into power spectrum data which is fed into the next layer of the next layer. Updating rule for the parameters of the correlational filter is derived using the back-propagation learning scheme. Experimental studies demonstrated that our feed-forward neural network produces superior performance for the classification problem with the patterns that have frequency information dominant property. Key-Words: Correlational operation, classification, Fourier transform, feature extraction, domain transformation, neural network.