Semantic multimedia analysis using knowledge and context

Abstract

The difficulty of semantic multimedia analysis can be attributed to the extended diversity in form and appearance exhibited by the majority of semantic concepts and the difficulty to express them using a finite number of patterns. In meeting this challenge there has been a scientific debate on whether the problem should be addressed from the perspective of using overwhelming amounts of training data to capture all possible instantiations of a concept, or from the perspective of using explicit knowledge about the concepts’ relations to infer their presence. In this thesis we address three problems of pattern recognition and propose solutions that combine the knowledge extracted implicitly from training data with the knowledge provided explicitly in structured form. First, we propose a BNs modeling approach that defines a conceptual space where both domain related evi- dence and evidence derived from content analysis can be jointly considered to support or disprove a hypothesis. The use of this space leads to sig- nificant gains in performance compared to analysis methods that can not handle combined knowledge. Then, we present an unsupervised method that exploits the collective nature of social media to automatically obtain large amounts of annotated image regions. By proving that the quality of the obtained samples can be almost as good as manually annotated images when working with large datasets, we significantly contribute towards scal- able object detection. Finally, we introduce a method that treats images, visual features and tags as the three observable variables of an aspect model and extracts a set of latent topics that incorporates the semantics of both visual and tag information space. By showing that the cross-modal depen- dencies of tagged images can be exploited to increase the semantic capacity of the resulting space, we advocate the use of all existing information facets in the semantic analysis of social media.