1 Image processing techniques for the blind removal of ghosts in archived film material Bruno Lopes, Joao Sequeira, Jean-Hugues Chenot, Lorcan Mac Manus and Anil Kokaram Abstract — This paper deals with blind ghost removal from archived film material. Unlike some other methods that work well for broadcast, the methods developed here cannot make use of deterministic signals placed in the video for channel charac- terisation. Instead, characterisation of the distorting mech- anism is achieved by examining the significant vertical edges present in the image. Three methods for restoring the dis- tortions in the image are then proposed and evaluated. Keywords —Ghosting, archived film material, echoes, over- shoot, blind equalisation of video I. Introduction T HE problem of echoes and overshoot in movie mate- rial is a well recognised one. The problem is perhaps most severe in broadcast media where multipath propaga- tion can add significant echoes and other undesired effects to the received picture. This whole family of undesired ef- fects is generally referred to as ghosting. The ghosting in received images can be a serious problem and various tech- niques have achieved a high success rate in dealing with it. The most successful methods have focussed on techniques that estimate the channel distortions, either by using a de- liberately transmitted Ghost Cancelling Reference (GCR) signal, or by exploiting other deterministic features of the television signal (for example, teletext) [1], [2]. There has been a recent resurgence of interest in the deghosting of images from the image processing commu- nity. Algorithms for the detection of objects in digital video streams have to isolate objects from the scene and cope with shadows and ghost images in the signal. These approaches typically use complex image processing tech- niques, such as image segmentation, for the blind identifica- tion and elimination of ghost images [3], [4], and deghost- ing of the images becomes an aspect of the overall object identification problem. In this paper, we are concerned with deghosting images that are already archived. This is an important problem, as television content providers strive to supply the enormous capacity of digital television by using previously broadcast material, which must be digitally enhanced prior to trans- mission. As the film has been previously archived, the recovery of deterministic information that could be used to identify the distorting mechanism is not possible. The Echoes and Overshoot Video material gently conceded by RTP and INA archives, Bruno Lopes and Joao Sequeira are with RTP, Av 5 de Outubro, 197 1000 Lisboa, Portugal. Jean-Hugues Chenot are with INA, Institut National de l’Audiovisuel, Departement de la Recherche, 94366 Bry sur Marne Cedex, France. Lorcan Mac Manus and Anil Kokaram are with the Department of Electronic and Electri- cal Engineering, University of Dublin, Trinity College, Dublin 2, Ire- land. Email: bruno.lopes@rtp.pt, Tel: +351 217 945 775 jhc@ina.fr, Tel: (33) (0) 1 49832009 methods presented in this paper are similar to those pre- sented by [5], in which a cross correlation method is used to estimate the delay and amplitude of the ghosted vertical edges in a noisy television picture. The structure of the rest of this paper is as follows. In the next section, a mathematical model is presented for ghosts, from which three separate correctional strategies are proposed. The paper concludes by presenting results from each of these strategies, and makes some final remarks on the area and the work. II. Mathematical Model for Echoes and Ghosting To eliminate the ghosts from the distorted image, it is first necessary to estimate the delay and amplitude of the ghost images so that the distorting mechanism may be modelled. This then provides a means to apply the inverse operation in order to restore the image. As pointed by [5], the problem of echoes occurs only at the vertical edges. It is proposed then that this distortion may be approximated as a process which operates in the progressive horizontal spatial direction of the original image. The most notice- able effect of this process is that it distorts step functions. The effects can therefore be most easily characterised by examining the edges present in the image and their im- mediate neighbourhoods, which should ideally behave like step functions. Since the source images are not available, the characterisation of the distortion of the edges must be performed without complete information on the original nature of the edges. To overcome this problem, the set of edges which are used to characterise the distortion pro- cess are limited to those which can be considered significant edges. For the purposes of this discussion, significant edges are those in which the luminance (grey level) transition is relatively large. This implies that a threshold must some- how be selected to allow the edge detection algorithm to only identify these significant edges. This threshold may be selected manually by examining a small, representative sample of images. Edge detection was performed using a Canny edge detec- tor. The edge detector detects all edges in the image, but as stated above, the distorting process operates only in the horizontal spatial domain. It is necessary then to eliminate the horizontal edges, which have not been distorted by this process, and leave the distorted vertical edges, so that the distortion may be properly characterised. The edge detec- tion process, and the removal of the unwanted horizontal edges is shown in Figure 1.