The application of classification wavelet networks to the recognition of transient signals

It has been demonstrated before that wavelets and neural networks can be successfully combined for the classification of transient signals. In the so-called classification wavelet network (CWN) the wavelet transform is applied in the first hidden layer of the network to extract compact features from input signals. This is followed by further layers to perform classification of the wavelet features. Since the definition of features is crucial to the success of a neural classifier, it may be advantageous if the definition of the features could be optimised. This paper demonstrates that the definition of wavelet-based features can be optimised as part of the training process of the CWN. The parameters that are optimised are the dilations and translations of the wavelet components. The mutual information criterion is employed initially to select a relatively small set of wavelets from the time-frequency grid. It is then demonstrated that the initial performance of the CWN can be improved by optimising the definition of the wavelet-based features. This technique is evaluated by application to the classification of seismic buffers