Wavelets and neural networks-based multiscale modeling with application to human face recognition

This dissertation developed new multiscale models for image processing. Multiscale features of images are generated by means of wavelets. The proposed models made few or no assumptions about the statistics of the multiscale process under investigation. The work was limited to the study of “white wavelets” and “linearity structure” assumptions in wavelets-based multiscale models. The proposed approach considered the wavelet coefficients as a second source of information rather than a white noise accounting for model uncertainties. Neural networks are used to get around the assumption of model linearity by learning the map between the nodes (e.g. pixels) residing at different levels (i.e. scales) of the tree (e.g. process). The nonlinear structure led to the development of a new multiscale technique for image segmentation that can be applied in a parallel scheme. Two metrics based on the first and the second order statistics of the model estimates were defined to measure the quality of the proposed multiscale models.; This dissertation also studied the feasibility of applying the proposed multiscale models to human face recognition. A subset of 20 faces from the Olivetti Research Ltd. (ORL) database were used to verify and validate the application of the proposed multiscale models to face recognition. Simulation study showed promising results in terms of recognition accuracy and sensitivity to illumination conditions.