Think organisms are the only ones with genomes? Researchers at the Israel Institute of Technology are sequencing the “video genome” to put an end to video piracy on the internet.
The technique works by detecting features that remain basically unchanged by typical color and resolution manipulations. Current methods rely on action recognition algorithms, which match video sequences by the movement they contain.
Think of three sample clips: the original lightsaber fight scene from Star Wars, a low quality video of the scene playing on TV and a home video of you and your brother reenacting it with plastic lightsabers. Action recognition algorithms would see all three clips as similar, but video genome analysis would only match the first and the second.
Brothers Alexander and Michael Bronstein and their advisor Ron Kimmel are able to make this distinction using gene sequence matching and alignment algorithms borrowed from the field of bioinformatics. Their research was posted on ArXiv.org on March 27.
“We realized that many problems that exist in the analysis of video match nicely to applications and problems that exist in the analysis of sequences,” Michael said.
The technique rests on the idea that changes like clipping and cropping of video are analogous to mutations in DNA. The “DNA” of video clips can be aligned the same way that biological DNA sequences can, using bioinformatics, even with the addition of commercials, deletion of scenes or changes to color or resolution.
“Looking at a very short piece of video, we are able to tell where it comes from, independently of the transformations it may undergo,” Alexander said.
For example, when a camcorder is used to capture the video from a movie screen, the camera may be shaking, the colors might be different, the resolution may vary, but the video DNA sequence would still be similar.
The video’s features are translated into a string of information, like a genome is read as a DNA nucleotide sequence. This video genome is made up of a group of features including boundaries and shapes, in the same way that search algorithms use a group of words to find similarities in text. These features don’t change during normal video manipulations.
“You can think of modifications a video can undergo as analogous to mutations,” Michael said. For example, an advertisement would be like an insertion mutation, and removal of content for rating would be a deletion.
The frequency of features in each frame is graphed and translated into a 64-bit binary word. When this information is played over time, the clip’s video genome can be compared to a database using bioinformatic analyses. The video genome takes up about one millionth the bandwidth of DVD quality video, and can be translated in real time.
The technology could potentially be used to detect pirated content on YouTube, or to match metadata, like subtitles, user-generated notes or comments, to any version of a video. Theoretically, thousands of hours of video could be processed in a matter of days, with greater than 99 percent matching accuracy.
Though the idea is promising, there are some drawbacks to the technology. While the video genome can be created in real time from video, matching the genome to entries in the database cannot. The team is currently working to simplify and speed up the database search process, so matches can be made in a fraction of a second.
Read More
http://www.wired.com/wiredscience/2010/04/video-genome/#ixzz0kEskrMQl
