Anomaly detection needs to learn one-class classifiers from normal instances in observation or feature spaces. In the Neyman-Pearson criterion, the design of one-class classifiers boils down to finding the minimal-vol...
详细信息
Anomaly detection needs to learn one-class classifiers from normal instances in observation or feature spaces. In the Neyman-Pearson criterion, the design of one-class classifiers boils down to finding the minimal-volume decision region subject to the error probability of normal instances no larger than a desired false alarm rate. The theoretical solution to this design problem is the probability density function (pdf) level set of normal instances. In low-dimensional feature spaces, by combining training samples with the convexity regularity on decision regions, the convexhull learning algorithm is a technique for designing one-class classifiers. In order to overcome its dimension limitation and the mismatch of convexity to the level sets of a multimodal pdf, this article considers the approach to replace the convexity by the connectivity to regularize decision regions. A fast graph-based maximum connected-component learningalgorithm is proposed to design one-class classifiers in high-dimensional feature spaces, which exploits the fast maximum connected-component search algorithm in a large-scale undirected graph. Moreover, for the application of sea-surface small target detection, the proposed algorithm combines ten-dimensional features to design feature-based detectors. Experimental results on the recognized radar database indicate the effectiveness of the proposed algorithm.
暂无评论