MULTIPLE CLASS MULTIPLE-INSTANCE LEARNING AND ITS APPLICATION TO IMAGE CATEGORIZATION

Abstract

We propose a Multiple Class Multiple-Instance (MCMI) learning approach and demonstrate its application to the problem of image categorization. Our method extends the binary Multiple-Instance learning approach for image categorization. Instead of constructing a set of binary classifiers (each trained to separate one category from the rest) and then making the final decision based on the winner of all the binary classifiers, our method directly allows the computation of a multi-class classifier by first projecting each training image onto a multi-class feature space and then simultaneously minimizing the multi-class objective function in a Support Vector Machine framework. The multi-class feature space is constructed based on the instance prototypes obtained by Multiple-Instance learning which treats an image as a set of instances with training labels being associated with images rather than instances. The experiment results on two challenging data sets demonstrate that our method achieved better classification accuracy and is less sensitive to the training sample size compared with traditional one-versus-the-rest binary MI classification methods.