Deriving kernels from generalized Dirichlet mixture models and applications

In the last few years hybrid generative discriminative approaches have received increasing attention and their capabilities have been demonstrated by several applications in different domains. Hybrid approaches allow the incorporation of prior knowledge about the nature of the data to classify. Past work on hybrid approaches has focused on Gaussian data, however, and less attention has been given to other kinds of non-Gaussian data which appear in many applications. In this article we introduce a class of generative kernels based on finite mixture models for non-Gaussian data classification. This particular class is based on the generalized Dirichlet distribution which have been shown to be effective to model this kind of data. We demonstrate the efficacy of the proposed framework on two challenging applications namely object detection and content-based image classification via the integration of color and spatial information.

► A new approach for the learning of generalized Dirichlet mixtures is proposed. ► Several kernels are derived from the generalized Dirichlet mixture. ► The proposed hybrid models are applied to object detection and image classification.