MPEG-2 Video Coding - Digital Video and Audio Broadcasting Technology - page 140

The matrices used in MPEG-4 AVC have a size of only 4x4 or 2x2 pix-
els, respectively. In the case of luminance, the transformation is performed
in two steps. In the first step, the original 4x4 pixel blocks are transformed
into the spectral domain by means of the matrix T. Following this, the
DCT coefficients of 16 blocks are again transformed by means of the Ha-
damard matrix H so that they can be compressed further (Fig. 7.37.)

DC coefficients Example:

16x16 luminance
macro block

Hadamard
Transform

Fig. 7.37. Hadamard transform of the DC coefficients in MPEG-4 AVC

In MPEG-2, it is either the matrices specified in the Standard which are
used, or they are specified by the encoder and modified and in each case
transmitted to the receiver in the sequence header at the beginning of a se-
quence. In addition, in MPEG-2, each coefficient is divided by the quan-
tizer scale factor which ultimately determines the actual data rate. MPEG-4
uses a set of 52 quantization matrices.
MPEG-4 also uses a deblocking filter (Fig. 7.38.) which is intended to
additionally suppress the visibility of blocking artefacts. This is also aided
by the smaller block size and the variable macroblock and slice size.

Reference
frames buffer

Deblocking
filter

Inverse
transform

Dequanti
zation

Motion
estimation

Entropy
coding

Quantization

Transform

Uncompressed
video

Compressed
video

Fig. 7.38. Deblocking filter in MPEG-4