Game Development Reference
In-Depth Information
The deblocking filtering operations on chroma component are the same as the
above operations in luminance component.
7.3 Sample Adaptive Offset
7.3.1 Overview of Sample Adaptive Offset
Although the deblocking filtering can reduce the blocking artifacts by filtering pixels
around block boundaries, the artifacts inside the blocks, e.g., ringing or blurring
artifacts are untouched. In order to reduce compression artifacts, a new and simple
in-loop filter, Sample Adaptive Offset (SAO), is widely discussed in latest video
coding standards, i.e., HEVC and AVS-2 (Fu et al.
2012
; Chen et al.
2013a
). SAO
aims to reduce the compression distortion not only in boundary pixels but also the
pixels inside blocks. In general, it is located after the deblocking filter. The key idea of
SAO is to reduce mean sample distortion by compensating an offset to each sample,
and the offset need transmit to the decoder. In order to trade-off the side information
(i.e., bits used for offsets) and coding efficiency, SAO first classifies the samples
into multiple categories with some classifiers, and then obtains an offset for each
category, which is applied to all the samples of the category. In the development
of video coding standards, many SAO methods or similar filters are proposed to
improve the coding efficiency.
In Mccann et al. (
2010
), an effective filter, extreme correction (EXC), was pro-
posed to HEVC. It first classifies the pixels into six categories by comparison of
current pixel value with upper, lower, left, and right neighbors (for nonboundary
pixels), and then for pixels in each category, the mean difference between the recon-
structed signal and the original signal is calculated as the transmitted offset. The
decoder can carry out the same operation to locate pixels of different categories.
Band Correct (BDC) proposed in Mccann et al. (
2010
), classifies the pixels sim-
ply according to their value ranges to avoids the comparison operations in EXC.
However, both EXC and BDC are both picture-based approaches, which obtain off-
sets by scanning the entire reconstructed picture, and then are applied to the image,
increasing the encoding latency significantly.
In order to reduce the encoding latency, Fu et al. (
2010
,
2011a
) proposed to
combine sequential stages of adaptive offset techniques into one stage and allow the
encoder to select only one classifier for each region adaptively. Another significant
simplification of SAO was proposed in Fu et al. (
2011b
), which derived offsets for
different regions aligning with largest coding unit (LCU) boundaries. Each region is
allowed to switch between band offset (BO) and edge offset (EO). In addition, the
only four 1D edge classification patterns are used instead of the 2D classification
patterns used in Fu et al. (
2011a
). In order to reduce the side information, only
offsets of partial bands are signaled to the decoder. Moreover, to further achieve low
latency of only one Largest Coding Unit (LCU), an LCU-based filter is proposed in
