Overview

Text which appears in digital videos can provide important supplemental information for indexing and retrieval. Although some work has been done on the extraction of text from images and isolated video frames very little work has considered the temporal aspects of video. In our previous work, we provided a methodology which identifies both scene and graphic text from low resolution video key-frames. We have found it necessary, however, to consider temporal changes to detect new information, to avoid processing redundant information, and to provide a means of integrating noisy results over multiple instances.

Text may span tens or even hundreds of frames, providing a tremendous amount of redundant information. By tracking moving text we can not only reduce the processing time by not applying all steps (detection, extraction , enhancement, OCR, etc) to every frame, but also maintain integrity by detecting when new information appears. For example, for ticker'' or scrolling'' text, new text continuously appears the screen so the system may need to process parts of subsequent frames to achieve a complete detection. Furthermore, we can make use of the temporal consistency to filter out false detections while having multiple instances to provide enhancement for OCR.

Researchers

          Huiping Li
          David Doermann

Publications

Automatic Text Tracking in Digital Videos(.ps) , in MMSP98
Automatic Text Extraction and Tracking in Digital Video, Lamp Technical Report

Demos

Demos for text tracking (initial text position is specified manually)

• Tracking conference video(.mpg)
• Text in athlete's jersey(.mpg)
• Return of Jedi(.mpg)
• Movie Credit in Complex Background(.mpg)

• Span camera across the wall(.mpg)
• Text zooming in(.mpg)
• Movie credits moving cross varying background(.mpg)
• Welcome to Lamp(.mpg)
• Star War(.mpg)