A Choquet Integral-based Meta-Classifier and its Applications on Multi-Class Image Classification

C. Campos and M. Mandal (Canada)

Keywords

Choquet Integral, fuzzy measure, MPEG7 Descriptors, entropy, mutual information

Abstract

In this paper, a novel information-based technique for predicting images types is introduced. In particular, we propose a framework of five Choquet integrals (i.e. one Choquet Integral for each image class) that are specialized to compute the global score of each image type. These global scores are obtained by evaluating a few selected MPEG-7 descriptors and provided by different individual classifiers. To compute the fuzzy measures associated with each Choquet Integral, we use an unsupervised method proposed in the literature [1] in which the concept of importance of attributes (in our case, the importance of the subsets of the individual classifiers) is replaced by that of information content in the subsets of classifiers. The parameters of the individual classifiers are adjusted with a training dataset of 500 images that are classified into five classes: animal, beach, city, bus, and view. The results obtained in this paper, shows that our proposed method obtains a higher classification accuracy compared to the results obtained for some popular methods cited in the literature.

Important Links:



Go Back