Page 60 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 2.6. Intraframe Coding 37

There are many variants of VQ [29]. Examples include adaptive VQ, clas-
si ed VQ, tree-structured VQ, product VQ (including gain=shape VQ, mean=
residual VQ, and interpolative=residual VQ), pyramid VQ, and nite-state VQ.
Theoretically, VQ is more eGcient than scalar quantization for both cor-
related and uncorrelated data [48]. Thus, the scalar quantizer in predictive,
transform, and subband coders can be replaced with a vector quantizer.
Vector quantization has a performance that rivals that of transform
coding. Although the decoder complexity is negligible (a lookup table), the
high complexity of the encoder and the high storage requirements of the
method still limit its use in practice. Like transform coding, VQ su6ers from
blocking artefacts at very low bit rates.

2.6.5 Second-Generation Coding

The coding methods discussed so far are generally known as waveform coding
methods. They operate on pels or blocks of pels based on statistical image
models. This classical view of the image coding problem has three main dis-
advantages. First, it puts more emphasis on the codeword assignment (using
information and coding theory) rather than on the extraction of representative
messages. Because the encoded messages (pels or blocks) are poorly repre-
sentative in the rst place, a saturation in compression is eventually reached
no matter how good is the codeword assignment. Second, the encoded entities
(pels or blocks) are consequences of the technical constraints in transforming
scenes into digital data, rather than being real entities. Finally, it does not
place enough emphasis on exploiting the properties of the HVS. E6orts to
utilize models of the HVS and to use more representative coding entities (real
objects) led to a new class of coding methods known as the second-generation
coding methods [49].
Second-generation methods can be grouped into two classes: local-operator-
based techniques and contour=texture-oriented techniques. Local-operator-
based techniques include pyramidal coding and anisotropic nonstationary
predictive coding, whereas the contour=texture-oriented techniques include
directional decomposition coding and segmented coding. Two commonly used
segmented coding methods are region-growing and split-and-merge. For a
detailed discussion of second-generation methods, the reader is referred to
Refs. 49, 50, 51.
Second-generation methods provide higher compression than waveform cod-
ing methods at the same reconstruction quality. They also do not su6er from
blocking and blurring artefacts at very low bit rates. However, the extraction
of real objects is both diGcult and computationally complex. In addition, such
methods su6er from unnatural contouring e6ects, which can make the details
seem arti cial.