Game Development Reference
In-Depth Information
efficient context modeling methods (Weinberger et al.
1996
;Wu
1997
;Xuetal.
2006
) for image coding.
In high-quality video coding, most of the bit budget is spent on transformed
prediction errors, called DCT coefficients. Consequently, how efficiently DCT coef-
ficients are entropy coded will significantly impact the coding efficiency of the whole
video coding. For the hybrid video coding systems, the statistical behaviors of DCT
coefficients are usually diverse. This diversity can be observed on different cod-
ing conditions, e.g., different video contents and different quantization step sizes.
In particular, DCT coefficients at different frequency subbands also show different
statistical behaviors. Context-based entropy coding is to exploit such diverse statis-
tical behaviors to achieve higher coding efficiency.
Observed domain knowledge, which is an abstraction of statistical behaviors,
is usually used to guide context modeling in context-based video entropy coding.
The entropy coding in MPEG-2 (MPEG2
1994
) is an example. MPEG-2 can use two
VLC tables for adaptively coding intra or nonintra DCT coefficients. Another kind of
well-known domain knowledge inDCT coefficients is, along the zig-zag path of DCT
blocks, nonzero coefficients show a statistical decreasing tendency in magnitude and
the run-length of successive zero coefficients shows a statistical increasing tendency.
This domain knowledge has guided the context modeling in CAVLC (Bjontegaard
and Lillevold
2002
) and CABAC (Marpe et al.
2003
) inAVC/H.264 (H264
2003
). For
example, CAVLC uses multiple contexts, each associated with a VLC table, to adapt
to such statistical tendencies, so that local statistical variations of DCT coefficients
even in one DCT block are exploited which brings further coding efficiency. In
CABAC, the transform coding encompasses coding methods of significant map,
last nonzero coefficient flag and coefficient levels. In the latest ITU-T and MPEG
video coding standard, referred to as High-Efficiency Video Coding (HEVC), the
transform coefficient coding encompasses the mode-dependent scanning patterns
for intramode coding, and new coding methods for the last significant coefficient,
multilevel significance maps, coefficient levels, and sign data hiding (Sullivan et al.
2012
; Sole et al.
2012
).
The following sections present two context-based entropy coding schemes for
coding transformcoefficients inAVS1/AVS2 video coding standards. One is Context-
based 2D Variable Length Coding (C2DVLC) (Wang et al.
2006
) for AVS1-P2 and
the other is Context-based Binary Arithmetic Coding (CBAC) for AVS1 (Zhang et al.
2009
) and its variant for AVS2. Both C2DVLC and CBAC adopt context modeling to
achieve higher coding efficiency. The context modeling is also inspired by the above-
mentioned domain knowledge, but is designed to maximally exploit the statistical
features of DCT coefficients under different transform sizes.
6.2 Context-Based 2D Variable Length Coding
In this section, C2DVLC for AVS1-P2 Jizhun profile is presented. First, we outline the
basic coding procedure of C2DVLC in Sect.
6.2.1
. Then, we describe how the context
