Digital Signal Processing Reference
In-Depth Information
the quantization step to change depending on the position of
the transform coefficient. Bigger steps, and more quantization,
are applied to the coefficients representing higher spatial
frequencies.
8 × 8 DCT coeff block
DC
Figure 15.2. Frequency-dependent DCT transform coef
cient quantization.
15.2.5 H.264 Pixel Interpolation
Motion vectors specify the location of a matching frame. If the
match is not close, coding of the residual will require more bits
than just coding the frame without any prediction. To assist in
matching, the motion vectors in MPEG-2 allow matching to
a pixel or half-pixel resolution. To perform half-pixel matching,
the linear interpolation is used
that is the pixel in-between is
calculated as the average of the four adjacent pixels side-to-side,
top and bottom. H.264 improves upon this in two ways. First, it
provides a better interpolation, executing the half-pixel interpo-
lation using the three adjacent pixels on each side. Second, it then
allows for a further quarter-pixel interpolation, using a linear
interpolation between the nearest half-pixels. Quarter pixel
resolution for matching enables a better predictive match to be
made, with smaller residuals, resulting in higher compression.
e
15.2.6 H.264 Variable Block Size Motion Estimation
In the MPEG-2 standard, motion estimation is always made
through matching of 16
16 macrocells. In H.264, block sizes can
vary from 16
4, enabling more accurate segmentation
of moving regions of the frame. For luma prediction, block sizes
include 16
16 to 4
4. These
can be used in any combination within a single macroblock. For
16, 16
8, 8
16, 8
8, 8
4, 4
8, and 4
