부산대학교 도서관

Sea-surface small target detection in high-resolution sea clutter is always an intractable problem. Feature-based detection in multidimensional feature spaces is recognized to be an effective way, and therein, learning algorithms with controllable false alarm rate play an important role. In this article, a weighted biased soft-margin support vector machine (WBSM-SVM) algorithm is proposed to design two-class classifiers with controllable false alarm rate in feature spaces and the induced feature-based detectors can accurately control false alarm rate and have excellent detection ability of small targets in sea clutter. The WBSM-SVM algorithm contains three innovations. First, special two-class SVM classifiers are used to lower the loss from the one-class $\nu $ -SVM classifiers by additional use of the training samples of the feature vector from simulated returns of typical targets plus measured sea clutter. Second, extremely unbalanced misclassification weight of penalty factors of the two classes and the Mahalanobis distance of the training sample vectors are introduced in the SVM to meet the demand of extremely unbalanced false alarm rate versus target missing probability. Third, a biased classification boundary is used to tune the false alarm rate to the expected one. The experimental results on the two recognized databases for sea-surface small target detection show that the WBSM-SVM-based detectors attain better detection performance than existing feature-based detectors.