In-Loop Filter - Advanced Video Coding Systems - page 120

Chen et al. ( 2013a ), which derives SAO parameters for each LCU and reuses the SAO
parameters in neighboring LCU according to rate-distortion cost. In next subsection,
a detailed SAO design in AVS-2 is introduced.
7.3.2 Sample Processing
Sample adaptive offset (SAO) is after the deblocking filter in AVS-2. It takes LCU
as the basic processing unit. BO and EO parameters are adaptively derived for each
LCU, and in order to save the bits consuming on SAO parameters, merge technique is
utilized, where a merge flag is employed to indicate whether the SAO parameters of
the current LCU are exact the same with its neighbors. When merge flag is enabled,
all the following SAO parameters are not signaled but inferred from neighbors.
The LCU-based SAO in AVS-2 assigns different offset values for each recon-
structed pixel in a LCU depending on the sample classification, and SAO parameters
are adapted from LCU to LCU. In every LCU, two types of classification scheme
are used: edge offset mode (EO) and band offset mode (BO). In edge offset mode,
each pixel in a LCU is classified into one of five categories based on the current pixel
and its neighbors specified by one of four predefined patterns shown in Fig. 7.6 .In
order to trade-off between complexity and coding efficiency, categories the four pre-
defined pixel classifications are all the 1D directional patterns: horizontal, vertical,
135 ⓦ diagonal, and 45 ⓦ diagonal. The notation “c” represents a current pixel to be
classified, and notation “a” and “b” represent two neighboring pixels. The classifi-
cation rules and the category for each pixel are summarized in Table 7.2 . The offset
values to be added to the reconstructed samples are signaled for each category except
the category 0 in which the edge offset is not used. In band offset mode, the whole
sample value range is divided into 32 bands, and offset values are derived for each
band but only four offset are signaled and transmitted to the decoder.
Edge Offset mode first classifies the pixels using four 1D directional patterns
as above introduced in Fig. 7.6 . According to these patterns, four EO classes are
specified, and only one EO class can be selected for each LCU. For a given EO
class, samples of current LCU are classified into one of the five categories, which are
based on the rules defined in Table 7.2 . For pixels in each category except category 0,
an offset is derived by calculating the mean of the difference of reconstructed pixel
Fig. 7.6 Four 1D directional EO patterns