Direkt zum Inhalt
Bannergrafik (AG-Freisleben)

Videana (BMWI)

Cloud-based Software Services for Content-based Search in Images and Videos 


Content-based search is of particular interest for large video and media archives and IP-TV providers and broadcasters. Searching for certain image or scene content is an essential daily activity in such organizations. With the rapid proliferation of multimedia content, there is an increasing need for high-quality image and video search based on the audiovisual scene content – and not only on subjective manual annotations/tags.

We have developed a software toolkit that enables a completely new way of semantic/content-based video search. Users can search for particular objects, events, locations, or persons in a video archive or database. For this purpose, the audiovisual content is automatically analyzed in a pre-processing step and all relevant information is saved: what is to see and what is to hear in a scene. This information serves as a basis for a high-quality search. 

The following content analysis tools are part of this software suite:

  • Semantic concept detection: A lexicon of audiovisual concepts can be defined that consists of, for example, 100 concepts. A concept can be any kind of object, event, location, setting or person. The software learns the appearance of concepts based on appropriate training data (video shots that show/do not show the concept) and can be used afterwards to automatically annotate video shots in a video database/archive with the probability that the concept is present in a shot. Although such annotations might not be perfect, they can be very useful to successfully explore a video archive.
  • Video OCR: This is a tool for localization, segmentation and recognition of super- imposed text in video frames. Superimposed text is often directly related to audiovisual scene content and gives useful hints about the semantics of a scene. However, superimposed text is normally of low quality due to lower video resolution and possibly complex image background. Hence, standard OCR systems are not well suited for video OCR. Our software automatically detects superimposed text in images and video frames and removes the image background in order to improve a subsequent OCR process.
  • …and much more: face detection (detection of frontal faces, optionally combined with face recognition), camera motion recognition (recognition of horizontal motion (pan), vertical motion (tilt), and zoom), speaker recognition (recognition of speakers in video sequences), audio segmentation (detection of speech, music, noisy or silent sequences), shot boundary detection (detection of cuts and gradual shot changes).

Since 2005, several components have achieved top results at  annual international evaluation contests (TRECVID), e.g., in the tasks of shot boundary detection, camera motion recognition and concept detection in videos. In addition, our group conducts research in the field of service-oriented architectures (SOA), high-performance and Cloud computing, which are also important topics for the compute-intensive field of video content analysis. 


Funding: Bundesministerium für Wirtschaft (BMWI), EXIST Transfer

Contact: Michael Rink, Markus Mühling


Zuletzt aktualisiert: 06.01.2014 · schmid2v

Fb. 12 - Mathematik und Informatik

Verteilte Systeme (AG Freisleben), Hans-Meerwein-Straße 6, D-35032 Marburg
Tel. +49 6421/28-21567, Fax +49 6421/28-21573, E-Mail: freisleb@informatik.uni-marburg.de

URL dieser Seite: https://www.uni-marburg.de/fb12/arbeitsgruppen/verteilte_systeme/forschung/pastproj/videana

Impressum | Datenschutz