Some of the most powerful schemes for scalable video coding are based on the so-called t+2D paradigm. In these schemes, temporal redundancy is first exploited through a motion-compensated multiresolution decomposition and the resulting temporal subband frames are then generally spatially decomposed with a wavelet transform. Thus, temporal and spatial scalability are achieved, in particular when also the motion information is properly managed between resolution levels. However, spatial wavelet transform may not be the most appropriate for exploiting the spatial redundancy of the detail subbands, since very often the power spectral density of these frames is not as concentrated at low frequencies as in the case of natural images. Recently, we have shown that orthonormal 4-band transforms can provide similar or even better results, especially for textured video sequences. In this paper, we elaborate on this idea and compare several M-band transforms for spatial decomposition of the temporal detail frames. Several M-band filter bank designs are given and lapped transforms are also considered. We show by simulation results that lapped transforms achieve a rate versus distortion performance that is comparable and sometimes better to that of the dyadic biorthogonal 9/7 wavelet transform, for a lower complexity.
The motion-compensated temporal filtering is an essential part of a scalable wavelet-based video coding scheme, which applies a temporal wavelet transform in the motion direction over the frames of a video sequence. The lifting structure of the temporal filter bank tackled in this paper involves a predict operator which makes use of two motion vector fields to bidirectionally predict frames from their neighbouring ones. We show in this paper that there exists an optimal
algorithm allowing to estimate jointly these two motion vector fields subject to an optimization criterion directly related to the coding of the detail subbands. We provide an iterative suboptimal form of this algorithm implementing this approach. We show that this
algorithm provides substantial gains in terms of PSNR, with the same complexity as a separate estimation of the two motion vector fields.