Sparse representations in multi-kernel dictionaries for in-situ classification of underwater objects
Date
2017
Authors
Hosseini, Somayeh, author
Pezeshki, Ali, advisor
Azimi-Sadjadi, Mahmood R., advisor
Chong, Edwin, committee member
Luo, Jie, committee member
Kirby, Michael, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
The performance of the kernel-based pattern classification algorithms depends highly on the selection of the kernel function and its parameters. Consequently in the recent years there has been a growing interest in machine learning algorithms to select kernel functions automatically from a predefined dictionary of kernels. In this work we develop a general mathematical framework for multi-kernel classification that makes use of sparse representation theory for automatically selecting the kernel functions and their parameters that best represent a set of training samples. We construct a dictionary of different kernel functions with different parametrizations. Using a sparse approximation algorithm, we represent the ideal score of each training sample as a sparse linear combination of the kernel functions in the dictionary evaluated at all training samples. Moreover, we incorporate the high-level operator's concepts into the learning by using the in-situ learning for the new unseen samples whose scores can not be represented suitably using the previously selected representative samples. Finally, we evaluate the viability of this method for in-situ classification of a database of underwater object images. Results are presented in terms of ROC curve, confusion matrix and correct classification rate measures.