Game Development Reference
In-Depth Information
Besides intra and interprediction, transform coding is another useful technique
to further remove the correlations among signals. It transforms the spatial image
data (image samples or prediction residual samples) into a different representation
in the transform domain, where coefficients have lower entropy so that they can be
coded more efficiently. There are lots of orthogonal transforms that can be used to
efficiently remove the data correlation. The Karhunen-Loeve Transform (KLT) is
the optimal transform to remove data correlation (Hotelling
1933
), but it depends
on the statistics of input data and has higher computation complexity. Thus, KLT is
difficult to be applied in real video compression systems. The block-based Discrete
Cosine Transform (DCT) (Ahmed et al.
1974
) is widely utilized in the hybrid video
coding framework, which has very similar performance with KLT (Chen and Pang
1993
) and can be calculated with low computation complexity by separating it into
two 1D transforms. The Discrete Wavelet Transform (DWT) (Mallat
1989a
,
b
)is
another efficient and compact signal representation method which decomposes a
signal into component wavelets. These wavelets have the great advantage of being
able to separate the fine details in a signal. It has been successfully applied in the still
image compression standard, JPEG-2000 (Christopoulos et al.
2000
). In addition,
transformcan concentrate the energy into a small number of significant coefficients in
low frequency bands, whichmakes it easier to be quantized based on the human visual
system [e.g., large quantization steps for insignificant high frequency coefficients
(Reader
2002
)].
Quantization technique can reduce the visual redundancy by representing a range
of values with a single quantumvalue. It could largely improve the compression ratio,
whereas it introduces the distortions into the reconstructed images. There are mainly
two kinds of quantization methods: vector quantization and scalar quantization. Vec-
tor quantization jointly quantizes a group of data to find a better representative vector
with the smallest rate distortion cost. Due to the complexity, vector quantization tech-
nique is not widely used in video coding standards. Scalar quantization maps a range
of signals into one value with low complexity by simply dividing a signal by a con-
stant, and then rounding to the nearest integer. Uniform quantization is a simple
quantization method which equally divides the entire range into some intervals and
selects the middle value of each interval as its quantized value. Considering that the
video signals after prediction usually has a centralized distribution, another quan-
tizer with a dead-zone is proposed to improve the rate distortion performance by
enlarging the quantization interval containing zero (Sullivan and Sun
2005
). Since
human visual system is more sensitive to the variation in low frequency bands, and
less sensitive to the changes in high frequency bands, this allows us to greatly reduce
the amount of information in the high frequency components by applying a large
quantization step. One example of the quantization matrix used for 8
8 DCT trans-
formed luminance blocks (recommended by JPEG) is shown in Eq.
1.1
, where larger
quantization steps are for higher frequency bands.
×
