Eigenvector decomposition of a multiresolution operator

Pattern recognition using a multiresolution representation was investigated. This was addressed as an eigenvalue problem. Eigenvector decomposition of a multiresolution operator enabled a low-pass pyramid representation to be expressed in terms of the 'eignpatterns' of the operator. The findings show that different image features 'emerge' and can be recognised at different levels of the multiresolution structure, i.e. at different resolutions. The level depends on the feature size. This work has implications in the design of a neural network for pattern recognition, namely that a network could 'learn' the eigenpattern of a multiresolution operator. Patterns recognition processing could proceed in a top-down, hierarchical manner, beginning at a level of coarse features and using information from lower levels in the multiresolution structure to guide the processing of finer detail.