Classifier Design by a Multi-Objective Genetic Algorithm Approach for GPR Automatic Target Detection

摘要

GPR is an electromagnetic remote sensing technique, used for detection of relatively small objects in high noise environments. Data inversion requires a fitting procedure of hyperbola signatures, which represent the target reflections, sometimes producing bad results due to high resolution of GPR images. The idea proposed in this paper consists of narrowing down the position of hyperbolas to small regions, using a machine learning approach. A Multi-Objective Genetic Approach (MOGA) is used to design a Radial Basis Function classifier. High order statistic cumulants are employed as features to this framework. Due to the complexity of the formulated problem, feature selection can be done in two ways: either by MOGA alone, or acting on a reduced subset obtained using a mutual information approach. The chosen classifier was tested on experimental data, the results outperforming the one presented in literature, or achieving similar results with models of much lower complexity.