Contributions to unsupervised and supervised learning with applications in digital image processing

Résumé

This Thesis covers a broad period of research activities with a commonthread: learning processes and its application to image processing. The twomain categories of learning algorithms, supervised and unsupervised, have beentouched across these years. The main body of initial works was devoted tounsupervised learning neural architectures, specially the Self Organizing Map.Our aim was to study its convergence properties from empirical and analyticalviewpoints.From the digital image processing point of view, we have focused on twobasic problems: Color Quantization and filter design. Both problems have beenaddressed from the context of Vector Quantization performed by CompetitiveNeural Networks. Processing of non-stationary data is an interesting paradigmthat has not been explored with Competitive Neural Networks. We have statesthe problem of Non-stationary Clustering and related Adaptive Vector Quantizationin the context of image sequence processing, where we naturally havea Frame Based Adaptive Vector Quantization. This approach deals with theproblem as a sequence of stationary almost-independent Clustering problems.We have also developed some new computational algorithms for Vector Quantizationdesign.The works on supervised learning have been sparsely distributed in time anddirection. First we worked on the use of Self Organizing Map for the independentmodeling of skin and no-skin color distributions for color based face localization.Second, we have collaborated in the realization of a supervised learning systemfor tissue segmentation in Magnetic Resonance Imaging data. Third, we haveworked on the development, implementation and experimentation with HighOrder Boltzmann Machines, which are a very different learning architecture.Finally, we have been working on the application of Sparse Bayesian Learningto a new kind of classification systems based on Dendritic Computing. This lastresearch line is an open research track at the time of writing this Thesis.