An Exploration of Multimodal Document Classification Strategies
Scott D. Chen · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2011
This thesis explores multimodal document classification algorithms in a unified framework. Classification algorithms are designed to exploit both text and image information, which proliferates in modern documents. We design meta-classification schemes that combine and integrate state-of-the-art text and image feature-extractors with state-of-the-art classifiers. Meta-classifiers fuse information across modalities that differ in nature and hence have more information on hand to make decisions. This thesis also discusses strategies that exploit correlations not only within a single modality but also among modalities. Techniques that exploit correlations within a modality include image meta-feature vector combination and latent Dirichlet allocation-based image meta-feature extraction. Another technique that exploits correlations between text and image cleans image with text information. Experiments on real-world databases from Wikipedia demonstrate the benefits of metaclassification for multimodal documents.