Options
Scene categorization using large margin Gaussian mixture models
Date Issued
01-12-2010
Author(s)
Abstract
For several categories of scenes, the representation of an image by a set of local feature vectors is suitable. Such sets of local feature vectors for images of a scene category are well modeled by generative approaches such as Gaussian mixture models (GMMs). For confusable categories, the discriminative training of generative models has been shown to give an improved classification performance. In this paper, we propose to use large margin GMMs (LMGMMs) for scene categorization task. The LMGMM uses large margin principles similar to support vector machine (SVM), for discriminative learning of GMM. The posterior probabilities estimated using LMGMMs are used as soft labels in building the posterior probability support vector machine (PPSVM) based classifier. The performance of LMGMM based classifier is compared with that of conventional GMM, variational Bayes GMM, SVM, PPSVM based classifiers.
Volume
1