ABSTRACT
Contents 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.2 TV Content Analysis for Harmful Content Detection . . . . . . . . . . . . . . . . . . . . . 58
3.2.1 Extracting Harmful Clues from a Single Modality . . . . . . . . . . . . . . . . . 59 3.2.1.1 Audio Analysis Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 3.2.1.2 Visual Analysis Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.2.1.3 Textual Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2.2 Combining Low-to Medium-Level Extracted Clues in Multimodal Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2.3 Higher Level Semantics Extraction: The Role of Ontologies . . . . . . . 67 3.3 Knowledge-Based Framework for Violence Identification in Movies . . . . . . 68
3.3.1 Preprocessing-Segmentation Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 3.3.2 Audiovisual Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3.2.1 Visual Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 3.3.2.2 Audio Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.3.3 Domain Ontology Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3.4 Inferencing Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.3.5 Implementation and Experimentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 3.3.6 Extensions-Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.1 Introduction Although controversial, television is probably the most common medium of information and entertainment. Everyone has access to TV content through a number of end user devices (TV sets, mobile phones, PCs) and over a number of different communication channels, but still limited control over the received content. Personalization services enhancing the overall viewers experience are nowadays made possible and offered by a number of media service providers. However little progress has been noted on the development of intelligent, efficient, human-like technological solutions for automatic identification and filtering of undesirable broadcasted TV content, which could further facilitate the protection of sensitive user groups (e.g., children). To better understand the underlying processes and provide a more intuitive description of the benefits and functionalities arising from such technologies, an example use case is presented involving Bob and Mary’s family-themselves and their two children, a daughter, Pat, who is still a kid, and a son, Tom, who is a teenager. This use case further illustrates how advanced TV content filtering services will operate. In this use case, Bob has just bought a brand new TV set with content filtering capabilities. He plugs it to the power source and turns it on. The profile settings screen appears. Bob creates four distinct user profiles for each one of his family members and browses through a harmful content hierarchy to select the corresponding content that should be filtered out for each one of them, including himself. As Pat is still a kid, all content classified as harmful should be filtered out; thus the top class of the hierarchy is selected. As Tom is a teenager, who likes action movies with fights and explosions and is old enough to watch partial nudity, the corresponding harmful content subclasses are deselected for him. However sexual actions, total nudity, and murders are disallowed. Mary does not like murders, blood splatter, and boxing; therefore, the corresponding classes are selected for her to be filtered out. Finally, Bob creates an empty profile for himself, as he wishes to view any type of received content. The next step is to filter out (i.e., skip or blur) the selected categories of received content for the corresponding user profile. For this purpose, the TV set (or set-top box) is accompanied by dedicated filtering software, that further allows for user profile storage and harmful TV content filtering on mobile phones and PCs, providing thus the desired protection from objectionable content for Bob and Mary’s family.
