Krishna, AG and Sreenivas, TV (2004) Shared Mixture GMM Classifier. In: International Conference on Signal Processing and Communications, SPCOM '04, 11-14 December, Bangalore, India, pp. 476-480.
Restricted to Registered users only
Download (844Kb) | Request a copy
In this paper, we propose the shared mixture GMM classifier. Mixtures of a GMM which represent areas in the feature space must have very less overlap across classes for best performance. If not, patterns of a class belonging to regions of overlap score high likelihoods with not only the mixture of its own class, but also that mixture of another class thereby making a high contribution to the likelihood score with respect to the other class model. We propose a computational method of determining if such mixtures exist and using an automatic method of determining such mixtures based on discriminability between classes, we propose choosing the mixtures for each GMM from the set of all mixturcs of all GMMs. Mixtures that have significant overlap with other classes get shared between the contending classcs. Weights are bascd on cluster-class membership matrix defined in the paper.We have compared the performance of this shared mixture GMM with a conventional GMM, for a 14 class music instrument recognition task. The new model performs competetively compared to the conventional GMM, and outperforms it when there is lesser training data.
|Item Type:||Conference Paper|
|Additional Information:||Copyright 2004 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.|
|Department/Centre:||Division of Electrical Sciences > Electrical Communication Engineering|
|Date Deposited:||22 Aug 2008|
|Last Modified:||19 Sep 2010 04:35|
Actions (login required)