International Core Journal of Engineering 2020-26 | Page 170
2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM)
Improve the Efficiency of Low Frequency Non-
Separable Secondary Transform Based on Implicit
Multiple Transform Selection
Zhu Hong Jucai Lin
R & D Center
Zhejiang Dahua Technology Co.,Ltd
Hangzhou, China
[email protected] R & D Center
Zhejiang Dahua Technology Co.,Ltd
Hangzhou, China
[email protected]
Dong Jiang Jun Yin
R & D Center
Zhejiang Dahua Technology Co.,Ltd
Hangzhou, China
[email protected] R & D Center
Zhejiang Dahua Technology Co.,Ltd
Hangzhou, China
[email protected]
Abstract—In order to effectively adapt to the dynamic
transformation characteristics of the residual, multiple
transform selection (MTS) was accepted in next generation
video coding standard. The MTS can be divided into explicit
MTS and implicit MTS, the difference between them was that
the primary transform pairs are different, and the primary
transform pair of implicit MTS can be determined uniquely by
the size of the input block. However, the MTS is a separable
transform, the compression capability is limited. So the Low
Frequency Non-Separable Secondary Transform (LFNST) was
adopted to perform the second transform in primary
transform coefficients. But LFNST does not separate the
difference between explicit MTS and implicit MTS, which
would have effect on performance of LFNST. In this paper,
when implicit MTS is enabled, we proposed that if primary
transform pair is a combination of DCT2 (Discrete Cosine
Transform type II) and DST7 (Discrete Sine Transform type
VII), replacing DST7 by DCT2. Two schemes were designed
and tested. The experimental results show that the proposed
two schemes achieved 0.09% and 0.16% luma BD-rate
reduction for all intra configurations respectively, and coding
running time complexity nearly constant. transform is adopted in H264 [1]. On the basis of H264,
HEVC (High Efficiency Video Coding) enlarges the size of
integer DCT2 transform, and also introduces DST7 (Discrete
Sine Transform type VII) transform into intra prediction 4*4
luma residual blocks [2].
The transforms were based on fixed image statistical
assumptions, both in the H264 and HEVC. However, the
prediction residuals have dynamic transformation characters
in actual image, and the simplicity of the transform kernel
limits the coding gain. In order to better adapt to the dynamic
transformation characteristics of the residual, Zhao et al.
proposed enhanced multiple transform (EMT) for video
coding, the experimental results show that the EMT
significantly improves the coding efficiency [4]. The scheme
defines three transform sets, each transform set contains two
transform bases in DST7, DST1 (Discrete Sine Transform
type II), DCT5 (Discrete Cosine Transform type V) and
DCT8, and the selection of the transform set is related to the
intra prediction mode. In 2018 Ljubljana meeting, EMT was
simplified and renamed MTS (Multiple transform selection),
which is applicable only to luma components [5]. Only two
transform bases, DST7 and DCT8 (Discrete Cosine
Transform type VIII) were retained, and the dependence of
selection of transform set on intra prediction mode was
eliminated. The combination of MTS and DCT2 constitutes
the main content of the primary transform has been adopted
in the VTM5.0 [6] reference software for developing the
next-generation video coding technologies.
Keywords—next generation video coding; implicit multiple
transform selection; low frequency non-separable secondary
transform
I. I NTRODUCTION
Transform is a very important part of the video coding
standard. The transform process can remove the spatial
correlation of data, and further achieve the purpose of
compressing data through quantization and entropy coding
[1-2]. In theory, KLT (Karhunen-Loeve Transform) [3] is the
optimal transform to remove the correlation between the
signals. However, KLT is difficult to implement because
without fast algorithms. Therefore, it is necessary to find a
transform that the performance is similar to KLT and easy to
implement. The DCT2 (Discrete Cosine Transform type II)
is an acceptable transform for KLT, exerting similar
performance with easier implement. The 4*4 integer DCT2
978-1-7281-4691-1/19/$31.00 ©2019 IEEE
DOI 10.1109/AIAM48774.2019.00037
The MTS contains implicit MTS and explicit MTS,
which cannot exist at the same time, explicit MTS is enabled
by default. The type of horizontal and vertical transform
defaults to DCT2, when explicit MTS is enabled, the explicit
MTS transform selection can have the following pairs:
(DCT8, DCT8), (DCT8, DST7), (DST7, DCT8), (DST7,
DST7). There is need to traverse all of transform pairs to
select the best transform. When implicit MTS is enabled, it is
judged whether to change the transform type to DST7
according to the size of the input block. Therefore, there are
148