Reconstruction of Severely Degraded Image Sequences

The AURORA project (AUtomated Restoration of ORiginal Film Archives) is an E.U. funded ACTS project which began in Septem-ber of 1996. The partners include three broadcasters and holders of archives, The Insitut National L'Audiovisuel (INA), The British Broadcasting Corporation, Radiotelevisao Portuguesa, two industrial companies , Snell and Wilcox, Societe Generale de Teleinformatic, and the signal processing groups of three academic institutions the Digital Media Institute , (Tampere, Finland) Delft University (The Netherlands) and Cam-bridge University Engineering Dept. (U.K). The project, coordinated by INA, has the sole purpose of designing new tools for video restora-tion/enhancement. This goal is motivated by the lack of a complete set of advanced manipulation tools which would otherwise allow the more complete exploitation of the archive holdings of many of the larger broadcasters. Furthermore, with the oncoming rise in Digital Video broadcasting a higher demand on quality and quantity of archive material is perceived ; hence the requirement for real time restoration devices is set to become more exacting. The project therefore considers the usual cornerstones of video restoration : noise reduction, missing data detection and reconstruction, reduction of image unsteadiness; as well as the associated software and hardware implementation issues. This paper concentrates on new developments at Cambridge University with respect to missing data reconstruction using probabilistic formulations. Missing data is a common impulsive degradation in archived lm. The lm material is abraded or occluded by foreign material caught in the projection mechanism thus yielding bright and dark ashes of light : blotches. The treatment of missing data in image sequences has traditionally been achieved by the use of spatio-temporal median lters. 1, 2] presented a series of alternatives which illustrate the need for a detector to control the operation of an interpol-ator which may take the form of a rank-order operation or employ a model based constraint on the spatio-temporal evolution of the image sequence. Within an interpolation framework, the spatio-temporal AR model was shown to be extremely accurate at reconstructing texture in missing regions as large as 20 20 ? Work funded under E.U. contract AC072, AURORA.