基于气氛语义的场景分类

Image categorization and annotation with abstract semantics beyond cognitive semantic level is a difficult problem in image understanding. On the other hand, atmospheres semantics of film frames is important to convey the message behind the story. For this abstract attribute, contextual information and domain knowledge besides the traditional content features are more helpful. The domain knowledge is included by the semantic ontology and network structures learnt from data. The contextual information includes film topic, recording time, and audience response. Based on all the metadata, Bayesian networks are employed to fuse metadata and a causal strength is proposed to account for the causation. Finally, we provide extensive experimental results to demonstrate the efficacy of the proposed approach.