Syntactic and Semantic Concepts in Audio-Visual Media

In recent years tremendous amounts of audio-visual media have become available to the public and increased the demand for efficient access and retrieval methods. The research community working in content-based media retrieval has mainly focused on specific media types, such as sports videos, news broadcasts, and commercials. A widely neglected media type is historic film provided by film archives and museums. In the context of historic film, film scientists and archivists have research questions and requirements that are novel for content-based retrieval. In this thesis, we investigate novel requirements for content-based retrieval of archive film stated by film experts. From the abstract requirements of film experts we first derive specific lowerand higher-level, syntactic and semantic concepts to be retrieved automatically. Next, we develop techniques for the automatic retrieval of these concepts from archive film. The investigated films are challenging for retrieval due to their sophisticated editing and their low material quality which results in numerous artifacts. The contribution of this thesis are novel techniques and investigations for the retrieval of syntactic and semantic concepts in archive film. We develop detectors for lower-level concepts such as black frames and intertitles and perform comprehensive investigations of shot boundary detection in archive film material. We further analyze higher-level concepts: We propose methods for the extraction of semantically coherent scenes and synchronous audio-visual montage sequences and investigate the retrieval of motion composition and visual composition. The developed techniques are successfully applied to archive and contemporary films and enable efficient access to the film material. Additionally the methods assist film experts in their investigations and enable them to gain novel insights into the films.