Global Plus Local Jointly Regularized Support Vector Data Description for Novelty Detection.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• One-class classification (OCC) is crucial for outlier detection when abnormal data is costly to acquire.
• Support vector data description (SVDD) is a popular OCC method, but its performance is sensitive to parameter selection.
• The trade-off parameter (C) in SVDD significantly impacts the sphere's center and volume, making its selection challenging.

Purpose of the Study:

• To develop a novel SVDD approach that overcomes the limitations of parameter sensitivity in traditional SVDD.
• To enhance outlier detection accuracy by incorporating both global and local image region information.
• To introduce a robust method for regularizing the trade-off parameter in SVDD.

Main Methods:

• Defined novel distance metrics in Gaussian kernel space relative to image regions.
• Designed two probability densities: one for the global region and one for the local region.
• Developed information quantity and information entropy for regularizing the trade-off parameter, leading to GL-SVDD.

Main Results:

• The proposed global plus local jointly regularized support vector data description (GL-SVDD) demonstrated encouraging performance.
• GL-SVDD effectively utilizes both global and local image information to penalize potential outliers.
• Experimental evaluations on UCI datasets and hyperspectral cherry fruit data confirmed the efficacy of GL-SVDD.

Conclusions:

• GL-SVDD offers a promising advancement in one-class classification and outlier detection.
• The joint regularization strategy effectively addresses the parameter selection challenge in SVDD.
• This method shows potential for applications where acquiring abnormal data is expensive.