Image annotation techniques based on feature selection for class-pairs

Abstract

Image annotation technique can be formulated as a multi-class classification problem, which can be solved by the ensemble of multiple class-pair classifiers. Support vector machine (SVM) classifiers based on optimal class-pair feature subsets from the multimedia content description interface (MPEG-7) standard are used as the class-pair classifiers. We use a binary-coded chromosome genetic algorithm (GA) to select optimal class-pair feature subsets, and a bi-coded chromosome GA to simultaneously select optimal class-pair feature subsets and corresponding optimal weight subsets, i.e. optimal class-pair weighted feature subsets. We consider two kinds of methods for class-pair feature selection: a common optimal (or weighted) feature subset is selected for all the class-pairs, and an individual optimal (or weighted) feature subset is selected for each class-pair respectively. Majority voting scheme is used to combine the class-pair SVM classifiers. The experiments are performed on two different image sets to validate the performance of our image annotation techniques.