MASTER THESIS Video Watermarking - Computer Graphics Group ...

frequency transform. The choice is made per macroblock. Further, these blocks
are transformed to remove spatial correlation inside the blocks.
Basically, the H.264 standard provides 4×4 block transform only but it has
been extended to 8×8 blocks. The transform is a very close integer approximation
to 2D-DCT transform with pretty much the same features and qualities.
Then, the transform coefficients are quantized (Q), i.e. divided by
quantization factors and rounded. This irreversible process typically discards less
important visual information while remaining a close approximation to the
original samples. After the quantization, many of the transform coefficients are
zero or have low amplitude, thus can be encoded with a small amount of data.
Figure 6: Encoding process scheme
The quantized coefficients are dequantized, inverse transformed and added
to the predicted samples to form a block of potentially reference picture.
Finally, the intra prediction modes or the motion vectors are combined with
the quantized transform coefficients and encoded using entropy coding (EC).
Entropy coding consists in representing more likely values by less amount of data
and vice versa.
The standard offers two entropy coding methods: Context-based Adaptive
Variable Length Coding (CAVLC) and Context-based Adaptive Binary Arithmetic
Coding (CABAC). CABAC is by 5-15% more effective [10] but much more
computationally expensive than CAVLC.
Entropy encoded data are enveloped together with header information as a
slice into a NAL unit.
17