Cedric Thebault

Scratch detection supported by coherency analysis of motion vector fields

In this paper, a solution for detecting scratches in old films and videos is presented. The focus is on automation and to enable fast, cost-effective processing of mass archives enabling a later reuse and repurposing. A common problem of line scratch detectors are the distinction of real scratches from strong vertical image features. Nearly all common solutions require a user interaction in the detection process. We propose a solution, which minimizes the false detection caused by strong vertical image features. Therefore, we utilize the fact that scratches are typically uncorrelated to the underlying image content and behavior. The detection is done by a structural analysis of the image, taking into account precise motion information and textures to separate scratches from real image content. After coherency analysis of spatial and temporal information, the false detected scratches are rejected. The proposed method allows a fast scratch detection and removal in an automatic environment without human interaction. Preliminary tests and results are presented and discussed.