On the performance bounds of motion-compensated deinterlacing

The terms 'proscan up-conversion' or 'deinterlacing' stand for algorithms for the conversion of interlaced image sequences into progressive format by interpolating the lines which are missing due to interlaced scanning. Although deinterlacing is already known to improve TV image quality, such techniques will become even more important in future. Firstly, standard interlaced image sequences will be increasingly processed and displayed on computer platforms and thus usually require progressive scanning and an increased temporal resolution. Secondly, the emerging market of light valve projection equipment demands high-quality proscan up-conversion since the image quality of these display techniques relies heavily on the light efficiency of non-interlaced scanning. By describing deinterlacing as a problem of linear prediction theory, this paper derives upper MSE performance bounds for these algorithms. Thus, a theoretical framework is provided which allows to analyse the effects of various parameters like motion vector inaccuracy, temporal integration and memory constraints on the efficiency of motion-compensated deinterlacing. It is, for example, shown that these algorithms can be very sensitive to motion Vector inaccuracies especially if the image sequence is recorded with very short exposure time. Finally, it is briefly demonstrated that the presented analysis can easily be extended to the case of field-and frame-rate up-conversion of interlaced (100 Hz') and non-interlaced image sequences. (C) 1997 Elsevier Science B.V.