Large Area Reconstruction

The previous chapters have considered the missing data problem in terms of relatively small sizes of distortion. The idea that motion correction plays an important role even with moderate sizes of corruption was also presented. This chapter focuses attention on the situations in which severe corruption occurs. In such situations a large proportion of the image can be lost. One example is the case of extreme physical abrasion, as in Figure 8.1. The problem also manifests due to loss of video signal in transmission. This is illustrated in the second frame of Figure 8.2, where loss of video signal has caused the video recording device (Sony Betacam Broadcast Standard VTR) to duplicate the last successfully recorded line until the signal resumed. The result is a swathe of duplicated lines (up to 50% of the vertical height) in every other field of video. The solutions presented in previous chapters are not well equipped to deal with this situation. In the case of the ‘local’ processes such as the two-step median and early 3DAR reconstruction (MBI-1, MBI-2) techniques, the missing area is now too big to be successfully assigned one motion vector. The joint motion estimation, detection and reconstruction technique can fare much better since it acknowledges the non-stationarity of the image and motion contents at the outset and so can implicitly estimate the separate moving objects that may be involved. However, that solution is still block based, and this limits its ability to define the edges of moving regions. Furthermore, it relies heavily on spatial motion information to achieve motion reconstruction. In this case of large corruption, it will take some time before the reliable motion information (at the extremities of the blotch) can propagate into the affected region. However, it is possible to deploy the joint solution presented in the previous chapter in this situation by making the necessary adjustments. These include allowing for occlusion as well as using a Pixel based motion reconstruction strategy instead of a block one. This would necessarily incur some computational overheads.