Image Annotation and Topic Extraction Using Super-Word Latent Dirichlet

This research presents a multi-domain solution that uses text and images to iteratively improve automated information extraction. Stage I uses local text surrounding an embedded image to provide clues that help rank-order possible image annotations. These annotations are forwarded to Stage II, where the image annotations from Stage I are used as highly-relevant super-words to improve extraction of topics. The model probabilities from the super-words in Stage II are forwarded to Stage III where they are used to refine the automated image annotation developed in Stage I. All stages demonstrate improvement over existing equivalent algorithms in the literature