Landmine detection using boosting classifiers with adaptive feature selection

Abstract

In order to solve the problem of landmine detection in Forward-Looking Ground Penetrating Virtual Aperture Radar (FLGPVAR), the AdaBoost classification with adaptive feature selection (AFS-AdaBoost) is proposed. The feature selection is added into the traditional AdaBoost, which can reduce the training error of weak classifiers and improve the generalization capability of a strong classifier. The feature selection is based on a wrapper model, whose cost function is the performance of the classifier. Considering landmine detection one-class classification problem, the false alarm rate with constant probability of detection is chosen to be the cost function, which ensures the detection performance of strong a classifier. Processing of a real dataset show that AFS-AdaBoost is applicable to the landmine detection in FLGPVAR. Compared with traditional AdaBoost, the detection performance and generalization capability of AFS-AdaBoost are significantly improved.