LGND: a new method for multi-class novelty detection

Abstract

Multi-class novelty detection is a crucial yet challenging aspect for recognition systems. Several methods have been presented, which either concatenate multiple classes into a large artificial super-class, or combine several independent classifiers of each known class, or utilize the results of multi-class classifiers. However, these methods ignore the correlation within each class, or cannot be elegantly formulated in a joint model. To overcome these limitations, we propose a new local and global novelty detection model (LGND). Different from the previous approaches, LGND incorporates the local correlation with the global regularization in a unified framework. This new optimization model boils down to a convex quadratic programming with guaranteed global optimum solution. Furthermore, comprehensive discussions, including the relationship between locality and globality, the discussion on the parameters in LGND and the connections to multi-class classification, are also presented. LGND opens up a new way for multi-class novelty detection from both local and global perspectives. Experimental results on Corel5k, Caltech-256, and five UCI data sets confirm that LGND outperforms or is at least comparable to state-of-the-art methods.