Media Compression - Scalable Continuous Media Streaming Systems - page 34

2.5.1.1 Requantization

The first way to transcode video to a lower bit-rate is by increasing the quantization step
size. One example is the Cascaded Pixel-Domain Transcoder (CPDT) [51], in which encoded
video is fully decoded back into its original pixel-domain representation and then re-encoded
using larger quantizers to reduce bit-rate. The computational complexity is high as a complete
decode-encode cycle, including the time-consuming motion estimation process, is needed.
Another approach proposed by Assuncao and Mohammed [43] reuses the original motion
vectors in re-encoding, and thus eliminates the need for motion estimation. Additionally, by
computing motion compensation in the DCT domain using the MC-DCT function proposed
by Chang and Messerschmitt [52], the IDCT and DCT operations can also be eliminated.
Their results showed that it can achieve results comparable to CPDT with significantly lower
computational complexity.

2.5.1.2 Spatial Downscaling

In spatial downscaling, we reduce the video's spatial resolution, e.g., from X by Y pixels to
X /2 by Y /2 pixels. Compared to requantization, spatial downscaling is far more complicated
because of two reasons. First, given four DCT macroblocks, we have to synthesize a downscaled
DCT macroblock. Second, as four macroblocks are combined into one macroblock, we cannot
simply reuse the original motion vector for the new combined macroblock.
Downscaling can be performed either in the pixel domain or in the DCT domain. In the pixel
domain approach, the four DCT blocks are decoded back to their original pixel representation,
downscaled by pixel averaging, and then re-encoded. In the DCT domain approach, four DCT
matrices can be downscaled to one DCT matrix without IDCT using the algorithm proposed
by Natarajan and Vasudev [53].
For the second problem, instead of performing motion estimation for the downscaled frames,
we can also reduce the computations by confining the search window to a few candidate motion
vectors [46, 48] or compute the new motion vector directly from the four original motion vectors
[45, 47].

2.5.2 Transcoder Design
As the goal of the transcoder is to adjust the transcoded video bit-rate to match the available
network bandwidth, we first consider the feasible operating ranges for requantization and
spatial downscaling, shown in Figure 2.8 for two different MPEG-1 CIF (352

× 288) video

streams Education and Onthestrip .
There are two observations. First, spatial downscaling can achieve a wider range of
transcoded bit-rate compared to requantization, especially at the lower bit-rates. Second, there
are regions where both techniques are feasible, and requantization achieves higher PSNR,
especially at the higher bit-rates.
These observations suggest that to maximize visual quality, we should use requantization at
higher bit-rates and adopt spatial downscaling at lower bit-rates when requantization cannot
be used. These two observations are incorporated into the integrated transcoder presented in
this chapter.