Game Development Reference
In-Depth Information
Chapter 7
In-Loop Filter
This chapter provides an introduction to in-loop filters in AVS-2. The first part
presents the characteristic of compression artifacts caused by block-based video
coding methods, the necessity of in-loop filtering to improve video coding efficiency
and the quality of compressed videos. In the following three parts, we describe the
three important in-loop filters, i.e., deblocking filter (DF), sample adaptive offset
(SAO), and adaptive loop filter (ALF), respectively. The last part concludes this
chapter.
7.1 Concepts of Compression Artifacts
Block-based transform (e.g., DCT) is widely adopted in hybrid video coding
framework. Every frame is divided into a group of nonoverlapped blocks with differ-
ent sizes, each of which is predicted from intra- or interframes and transformed into
the frequency domain using block-based transforms. For each transformed block, the
coefficients are quantized independently and then compressed into a binary stream
via entropy coding. However, at low bitrate, the block-based video codecs usually
suffer from visually annoying compression artifacts, e.g., blocking and ringing arti-
facts. Figure
7.1
a shows a typical compression artifact example of one frame coded
by AVS-2 at low bitrate, blocking artifacts on the face and ringing artifacts around
the shoulder of the woman being obvious.
The above compression artifacts are mainly generated from two sources. The
most significant one is the independent coarse quantization of block transform coef-
ficients, which makes only few quantized coefficients left in one block and causes
the discontinuousness near block boundaries and image edge areas. The other source
of the compression artifacts is motion compensation prediction, which may copy
blocks from different positions of previous reconstructed frames. On one hand, the
artifacts in compensation blocks propagate to the current frame; on the other hand,
the block discontinuousness is aggravated when the adjacent blocks are copied from
different position or different reference frames.
