Digital Signal Processing Reference
In-Depth Information
15.2.2 H.264 Support for 4
4 Integer DCT
4 DCT, each transform coefficient has a contribution
from 16 pixels, compared to the 8
In a 4
8 DCT, where each coefficient
has a contribution from 64 pixels. Therefore the 4
4 DCT has
less computations per pixel.
One of the key elements in H.264 is the ability to compress
videousinga4
4 pixel blocks. In addition, many
of the DCT coefficients and operations are optimized for digital
arithmetic. Since the DCT output coefficients are to be quan-
tized afterwards, the precision of the DCT calculation can be
reduced. For example, coefficients which are close to 0.5 are
approximated to 0.5. This does not require a multiplier, but can
be implemented with a right-shift operation. Furthermore, some
divides can be eliminated by pre-multiplying the input data by
factors of two, which can be done with left shifts. These steps can
allow all coefficients to be simple integers. The net result is
a major simplification in the DCT, which is where a significant
portion of the processing takes place.
The same simplification process is applied to the 8
4DCTon4
8DCT
used in H.264. However, not all of the multiplies and divides can
be eliminated in the 8
8 DCT, due to its greater dynamic
range.
The simplifications are of particular help in parallel hardware
implementations, such as FPGAs and ASSPs, where the multiplier
elimination can save significant circuit area.
15.2.3 H.264 Logarithmic Quantization
The quantization in the earlier MPEG standards was linear. In
H.264 the Mquant parameter is replaced by the QP (quantization
parameter), which has a value from 0 to 51. However, the effect of
QP is not linear. Each increase of six in the QP value results in
a doubling of the quantization step size. The logarithmic scaling
provides finer step sizes for small QP, and coarser step sizes for
large QP. The result is a greater dynamic range for scaling, with
the same number of bits to represent the QP parameter in the
quantization process.
15.2.4 H.264 Frequency Dependent Quantization
Human vision is less sensitive to artifacts when there are
complex scenes, which tend to have high spatial frequencies.
Previously, the quantization applied to the DCT transform coef-
ficients has been uniform, meaning all coefficients use the same
quantization rules. Frequency-dependent quantization allows
