A wavelet network based classifier

This paper studies a wavelet network based classifier and introduces Orthogonal Least Squares method to quickly fulfill the structure and parameter identifications for dealing with the limitation of wavelet network in multidimensionality. First, we build up an initial multidimensional wavelet network with wavelet frame theory. Next, Recursive Orthogonal Least Squares is used to recursively orthogonalize the wavelets instead of batch Orthogonal Least Squares. After the orthogonalization, calculate the contribution of each wavelet to the output energy, and eliminate unimportant wavelets with backward elimination approach so as to establish an effective and refined wavelet network. Finally, numerical experiments are given to demonstrate the effectiveness of wavelet network based classifier.