Cryptography Reference
In-Depth Information
Distortion
λ
λ
2
λ
3
R3
R1
R2
Bit Rate
Fig. 3.12.
The optimal Lagrange parameter at different optimization points.
quality layer are coded by the non-scalable CABAC in AVC/H.264 [10],
J = Distortion + λ(BitRate) .
(3.5)
According to the theory of Lagrange optimization, the optimal Lagrange
parameter (i.e., the λ factor) varies with the optimization point as illustrated
in Fig. 3.12. Thus, if the criterion for mode selection is to minimize the La-
grange cost as defined in Eq. (3.5), different motion vectors and prediction
residues could be generated for different layers. Such flexibility allows the
motion vector and prediction residue be refined as the bit rate increases.
Fine Granularity Scalability
To achieve FGS, the non-scalable CABAC is replaced by an embedded cyclical
block coding [13]. By using embedded coding, the bit-stream of a quality layer
can be arbitrarily truncated for FGS.
In Cyclical Block Coding [13], each quality layer is coded using two passes.
These are the Significant Pass and the Refinement Pass. The significant pass
first encodes the insignificant coe
cients that have zero values in the sub-
ordinate layers. The refinement pass then refines the remaining coe
cients
within range 1 to +1. During the significance pass, the transform blocks are
coded in a cyclical and block-interleaved manner. By contrast, the refinement
coe
cients are coded in a subband by subband fashion [13].
Fig. 3.13 gives an example of cyclical block coding [13]. For simplicity, we
assume all the coe
cients are to be coded in the significant pass. As shown,
each coding cycle in a block includes an EOB symbol, a Run index and a non-
zero quantization level. The EOB symbol is coded first for indicating whether
there are non-zero coe
cients to be coded in a cycle. In addition, the Run
index represented by several significance bits is used to locate the non-zero
