Media Compression - Scalable Continuous Media Streaming Systems - page 37

buffer for the macroblock, which is then extracted using MC-DCT [52] and added to the next
predicted frame to compensate for the quantization error.
For spatial downscaling, as error is introduced in both quantization and DCT operations,
we need to perform both dequantization and inverse DCT to compute the aggregate errors. We
can then incorporate these errors into the residual errors in the motion compensation process
to prevent drifting.

2.5.3.2 Reconstruction of the Downscaled Motion Vectors

In spatial downscaling, instead of performing motion estimation again on the downscaled
frames, which is computationally expensive, we can use the AMVR algorithm proposed by
Bo et al. [45] to reconstruct the motion vectors for the downscaled frame. In AMVR, the
downscaled motion vector, denoted by MV , is computed from

4

i = 1

m v i A i

1
2

MV = (2.1)

i = 1

4

A i

×

where mv i is the original motion vector of the macroblocks i in the original N N video, and

A i is the activity measurement associated with macroblock i . For simplicity, A i is taken to be
the number of non-zero AC coefficients in the macroblock i which can be readily obtained
when the coefficients are parsed [45].

2.5.3.3 Quantizer Regulation

In order to match the output video bit-rate to the target bit-rate, a quantizer regulator similar
to the quantization control scheme in the Test Model 5 (TM5) [54] is incorporated to control
the quantizer scale in each macroblock.
Instead of using the original TM5 quantization control scheme directly in requantization,
we need to modify the quantizer regulation scheme so that the new quantizer scale will not be
smaller than the original one. Specifically, let q be the quantizer scale computed using the TM5
algorithm. Then the new quantizer scale, denoted by Q , is obtained from max q , Q

, where
Q is the original quantizer scale. This modification reduces bit-rate consumption as using a
quantizer smaller than the original one cannot improve visual quality and yet increases the
number of bits required to encode the quantized coefficient. Note that the same modification is
not required in spatial downscaling because DCT coefficients of the downscaled macroblocks
are reconstructed during downscaling.

{ }

2.5.4 Experimental Results
To evaluate experimentally the performance of the presented integrated transcoder, we imple-
mented the transcoder depicted in Figure 2.9 in software running under the Microsoft Windows
platform. We tested the transcoder using a variety of different videos. The results of two video
segments, referred to as Onthestrip and Education , are presented in this section. Both segments
are originally MPEG-1 encoded in CIF resolution with a frame-rate of 30 fps at 1.15 Mbps.