|Thesis abstract: |
In the last few years, with the rapid proliferation of inexpensive hardware devices and the increasing popularity of the Internet and social media, we have witnessed an exponential growth in the creation and redistribution of digital content. However, a significant fraction of multimedia content available online is not originally acquired by contents creators. Conversely, it is created by editing and reusing already existing material. This process is encouraged by the ever increasing availability of powerful and user-friendly editing software, which allows anyone to easily modify multimedia objects. Hence, images, videos and audio tracks can be seen as objects that evolve over time similar to organisms in biology, and each new version may, in turn, generate other versions. This evolution due to editing operations and alterations gives birth to near-duplicate objects, i.e., similar but not identical copies of the original files.
The detection and analysis of near-duplicate multimedia objects have received great attention from researchers in the last years under the name of multimedia phylogeny. More specifically, multimedia phylogeny investigates the history and evolutionary process of digital objects and includes finding the causal and ancestral document relationships, source of modifications, and the order and transformations that originally created the set of near-duplicates. This enables to shed very interesting insights on the way content is created and reused. As an example, once a pool of near-duplicates is isolated, it is possible to identify which object is the original one that gave birth to the near-duplicate copies. Moreover it is possible to infer the generative structure behind near-duplicates creation, in order to establish a relationship between all the near-duplicates pairs. These prove to be key-steps in many authenticity- and forensics-related applications.
In this thesis, we aim to develop a comprehensive set of multimedia phylogeny strategies for studying near-duplicates relationships, especially targeting video content. Indeed, video content analysis proves to be a challenging scenario still not fully developed in the literature. As a matter of fact, while it is straightforward to find exact duplicates in a set of available videos, detecting near-duplicates proves to be much more difficult. In fact, some operations (such as coding) are typically more aggressive on videos than on still images. These may introduce high levels of distortion on the media and prevent many systems from detecting near-duplicate copies. Moreover, in video analysis, the temporal domain must be explicitly taken into account. This increases the number of possible transformations that can be applied to a video with respect to still images (e.g., frame-rate conversion, time clipping, etc.).
The outcome of this thesis may have a strong impact in many applications such as:
i) security: the transformation history of a set of documents can describe the flow of content distribution;
ii) forensics: the original document can be identified out of a near-duplicate set to perform document forensics analysis;
iii) copyright enforcement: traitor tracing without watermarking or fingerprinting method can be devised;
iv) news tracking services: near-duplicate relationships can explain key elements about the opinion forming process across time and space.