Discriminative Multi-Modal LDA for Simultaneous Image Classification and Annotation

Abstract

With the help of powerful topic modeling techniques, image classification and annotation has now emerged with many meaningful applications in our real world. We propose a novel supervised topic model to automatically learn the image labels and annotation words. Based on previous work, we introduce the category information into the regression between the two modalities of LDA to help narrow down the set of feasible annotation words, and to be able to make a classification for testing images simultaneously. Our experimental results show the out-performance of our model over the previous topic regression multi-modal LDA for image annotation, with similar running time

and the competitive performance compared to recent LDA-based work for image classification.