DIVAS multimedia search engine finds content using digital ‘fingerprints’
By Darren Quick
April 14, 2010
Text-based searches might be great for hunting down relevant chunks of text-based information, but searches for multimedia content can be a little more hit and miss. Searches that rely on manually assigned metadata and often misleading titles can return off topic results, while searches that require the unpacking of compressed data can slow up the search. DIVAS is a new multimedia search engine that addresses these problems by using digital "fingerprints" that, according to its developers, return more reliable results.
Developed by researchers at the Franhofer Institute, DIVAS - short for Direct Video & Audio content Search engine - can be used to detect similarities between different video or audio contents as well as for topic-related data searches, for instance to find songs of a certain genre. The software can handle audio and video formats such as MP3, AAC, H.264 and MPEG-2.
It works its magic by examining the digital fingerprints of multimedia files. These fingerprints are stored in MPEG-7 format, the ISO standard for multimedia data. Fingerprints in music files can provide information on its tempo, genre or the proportion of rhythmic instruments, while those used in videos may contain information on scene changes, camera movements and the brightness of the image. In contrast to manually captured metadata, automatically generated fingerprints are never ambiguous.
In many archives – for example those of television broadcasters – such fingerprints are generated when the file is initially stored. This differs from the Internet, where fingerprints have to be recreated with every search. For this purpose, it was previously necessary to decompress compressed files. But that is no longer the case with DIVAS, which can find the file you are looking for without needing to first unpack the entire media archive.
The Franhofer researchers say the search engine is suitable not only for searches on the Internet and in archives, but also for monitoring TV programs, for example to check whether a contractually agreed advertisement has been broadcast. As the software can take fingerprints of compressed files, it is faster than comparable search engines.
“Our series of tests with MP3 files showed that search times can be halved,” reports Prof. Gerald Schuller of the Fraunhofer Institute for Digital Media Technology IDMT in Ilmenau. Together with Matthias Gruhne he was responsible for developing the audio component of DIVAS, which covers the extraction of fingerprints and the subsequent search. Both of these function even with music files that are recorded from a loudspeaker using a cell phone.
“Even if the sound has been severely distorted, our methods make it possible to clearly identify and classify the song,” explains Schuller.
DIVAS was the brainchild of an EU-wide joint project in which seven partners from seven different countries took part. They included Belgian television broadcaster BETV, which plans to use the multimedia search engine in its own archive.
There’s no word on when DIVAS will hit the web, but since the information provided by Franhofer indicates the technology relies on the structure of the multimedia content to perform its searches the question still remains as to how it can accurately identify the nature of the content. While analyzing the file structure might work pretty well for identifying the genre of music files, the same can’t be said for video content. Still, the proof of the pudding is in the tasting, so we’ll reserve judgment until we can give it a try.