A Kernel Machine Framework for Feature Optimization in Multi-frequency Sonar Imagery

摘要

The purpose of this research is to optimize the extraction of classification features. This includes the optimal adjustment of parameters used to compute features as well as an objective and quantitative method to assist in choosing a priori data collection parameters (e.g., the insonification frequencies of a multi-frequency sonar). To accomplish this, a kernel machine is employed and implemented with the kernel matching pursuits (KMP) algorithm. The KMP algorithm is computationally efficient, allows the use of arbitrary kernel mappings, and facilitates the development of a technique to quantify discriminating power as a function of each feature. A method for feature optimization is then presented and evaluated on simulated and experimental data. The experimental data is derived from low-resolution, multi-frequency sonar and consists of a large feature space relative to the available training data. The proposed method successfully optimizes the feature extraction parameters and identifies the (much smaller) subset of features actually providing the discriminating capability.