Markov random field based fusion for supervised and semi-supervised multi-modal image classification

摘要

In recent years, there has been a massive explosion of multimedia content on the web, multi-modal examples such as images associated with tags can be easily accessed from social website such as Flickr. In this paper, we consider two classification tasks: supervised and semi-supervised multi-modal image classification, to take advantage of the increasing multi-modal examples on the web. We first propose a Markov random field (MRF) based fusion method: discriminative probabilistic graphical fusion (DPGF) for the supervised multi-modal image classification, which can make use of the associated tags to enhance the classification performance. Based on DPGF, we then propose a three-step learning procedure: DPGF+RLS+SVM, for the semi-supervised multi-modal image classification, which uses both the labeled and unlabeled examples for training. Experimental results on two datasets: PASCAL VOC'07 and MIR Flickr, show that our methods can well exploit the multi-modal data and unlabeled examples, and they also outperform previous state-of-the-art methods in both two multi-modal image classification. Finally we consider the weakly supervised condition where class labels are from image tags which are noisy. Our semi-supervised approach also improves the classification performance in this case.