| This thesis proposes novel variations of Sparse Representation techniques and shows successful applications to a variety of fields such as Automatic Speech Recognition (ASR) denoising, Face Recognition and Underwater Image Classification. A section of the thesis makes new algorithmic contributions in Group Regularization which are able to better handle collinear dictionaries. A new ASR front-end is introduced, and applies these algorithms to feature denoising in ASR pipeline. The thesis also explores effective ways for dictionary partitioning for improved speech recognition results over a range of baselines. A new method for combining predictions from different feature windows is also explored. In addition to the denoising section, this thesis also proposes new methods for Sparse Representation Classification (SRC) which better couples the regularization and decision making step. The effectiveness of these new methods are shown in both Face Recognition and Underwater Image Classification. Since all these techniques are domain independent with the right feature extraction procedure and training set, they show great promise to be applied to a gamut of other different areas. |
