Detection of video sequences using compact signatures

Abstract

<jats:p>Digital representations are widely used for audiovisual content, enabling the creation of large online repositories of video, allowing access such as video on demand. However, the ease of copying and distribution of digital video makes piracy a growing concern for content owners. We investigate methods for identifying coderivative video content---that is, video clips that are derived from the same original source. By using dynamic programming to identify regions of similarity in video signatures, it is possible to efficiently and accurately identify coderivatives, even when these regions constitute only a small section of the clip being searched. We propose four new methods for producing compact video signatures, based on the way in which the video changes over time. The intuition is that such properties are likely to be preserved even when the video is badly degraded. We demonstrate that these signatures are insensitive to dramatic changes in video bitrate and resolution, two parameters that are often altered when reencoding. In the presence of mild degradations, our methods can accurately identify copies of clips that are as short as 5 s within a dataset 140 min long. These methods are much faster than previously proposed techniques; using a more compact signature, this query can be completed in a few milliseconds.</jats:p>