Normalization of Feature Distributions for Linear-Quadratic Fusion in Landmine Detection Using GPR

Successful detection of antipersonnel landmines often requires decision making on the basis of more than one decisive feature. The fusion of the available features should be done in a statistically optimal way. Theoretically this might be achieved by Bayesian-based criteria. However application of Bayes´ criteria in general form is computationally difficult and may lead to overtraining of classification algorithms. To avoid this situation linear and quadratic fusion rules are often used in practice. However these rules produce optimal results only when the decisive features are normally distributed, which is not always guaranteed in practice. To improve the performance of the linear and quadratic detectors we suggest a normalization of the decisive features by means of Johnson´s transform prior to the fusion. We describe the normalization algorithm and demonstrate that it improves performance of the linear and quadratic classifiers. The performance is judged in terms of the ROC curves