International Core Journal of Engineering 2020-26 | Page 171

four possible transform pairs of primary transform: (DCT2,
DCT2), (DCT2, DST7), (DST7, DCT2), (DST7, DST7). The
type of transform pair is uniquely determined according to
the size of the current block, and there is no need to traverse
all of transform pairs to select the best transform. Implicit
MTS is cut-down version of explicit MTS, these two modes
cannot coexist at the same time. When implicit MTS is
enabled, type of horizontal and vertical transform can be
uniquely determined using the size of the input block, DCT2
is used for the long side and DST7 for the short side.
transform both to DCT2. It is judged whether to change the
transform type to DST7 according to the size of the input
block. When the conditions of the length of width or height
are satisfied with greater or equal to 4 and less than or equal
to 16, select DST7 as the corresponding type of horizontal or
vertical transform. After the primary transform is completed,
according to the value of LFNST index, determine whether
the LFNST is enabled or disabled. There are three states of
LFNST: LFNST0, LFNST1 and LFNST2. Where LFNST0
means LFNST disabled, LFNST1 means LFNST enabled
and the transform kernel selects the first one in the transform
set that selected by intra prediction mode, LFNST2 means
LFNST enabled and the transform kernel selects the second
one in the transform set that selected by intra prediction
mode. Choose the best transform between 3 kinds of LFNST
options, and the coefficients of the best transform are further
compressed by quantization and entropy coding.
Implicit MTS can be classified into two cases: ISP (Intra
sub-partitions) mode or turn on implicit MTS switch [7-8].
ISP mode is an intra-block partition mode , which divides the
intra luma prediction block into 2/4 sub-regions according to
the size of vertical or horizontal, each sub-region has a
minimum of 16 sampling points, and the minimum CU
allowed to use ISP mode is 4*8/8*4. In the encoder, the sub-
regions are sequentially encoded in order from top to bottom
(left to right). After the coding of the previous sub-regions
was completed, the inverse transform and inverse
quantization were performed, and the predicted pixels are
added to reconstruct the pixels so that the next sub-regions
can be predicted. All sub-regions use the same intra-
prediction mode. The other case is to turn on the implicit
MTS switch and the Explicit MTS switch to turn off in the
code.
Although the MTS considers the dynamic transformation
characteristics of the residual, the MTS is a separable
transform and the compression capability is limited. In 2016,
a Non-Separable Secondary Transforms (NSST) was
proposed and implemented on top of HEVC [9]. The
proposed method performs small size non-separable
transforms on lower frequency transform coefficients of
primary transform, so the transform efficiency was further
improved and greatly enhances the overall performance of
encoder. In 2019 Geneva meeting, LG Electronics came out
Low Frequency Non-Separable Secondary Transform
(LFNST) and has been adopted in VTM5.0. On the basis of
NSST, the proposal proposed only keep the low-frequency
coefficients of the secondary transforms and the high-
frequency coefficients are zeroed, which greatly reduces the
storage space and the amount of calculation. [10]. There are
four transformation sets and each transform set contains two
non-separable transform matrices (kernels) in LFNST.
Among them, the selection of the transform set is related to
the intra prediction mode, and the selection of the transform
kernel is related to the LFNST index.
Conclusions can be drawn from previous related research
as follows: the coding efficiency of LFNST depends on the
LFNST core, and the LFNST kernel is trained by the main
transform coefficients when Explicit MTS is turned on [9-
10]. Therefore, the LFNST kernels are more suitable for the
primary transform types when the explicit MTS is enabled,
but these types do not contain the combination of DCT2 and
DST7 when the implicit MTS is enabled. So that implicit
MTS does not perform very efficiently for LFNST compared
with explicit MTS.
Fig. 1. The combination of implicit MTS and LFNST in VTM5.0.
III. P ROPOSED COMBINATION OF IMPLICIT MTS AND LFNST
According to the above analysis, this paper puts forward
two improving schemes for the implicit MTS. The main
thought of these schemes is: remove the type of transform
pairs that may affect the performance of the LFNST and turn
them into the transform pair suitable for LFNST. That is
when the primary transform pair type is a combination of
DCT2 and DST7, the DST7 is changed to DCT2. Two
experimental plans were designed as shown in Fig. 2 and Fig.
3.
Implicit MTS is simpler than explicit MTS, the
correlation research are not enough. In addition, the
performance of LFNST did not take into account the
difference between implicit MTS and explicit MTS.
Therefore in this paper, the influence of the difference in
implicit MTS and explicit MTS on the performance of
LFNST was studied, and puts forward some improvement
schemes. Experimental results shown that encoding
efficiency could be improved and coding complexity was
almost unaffected
The first experimental plan proposes as long as the
primary transform pair is a combination of DCT2 and DST7,
the DST7 is changed to DCT2, and then does the additional
processing. The second experimental plan proposes that
when LFNST index was 0, the LFNST was disabled, implicit
MTS does not make any changes regardless of the type
combination of the main transform pair, when LFNST index
was greater than zero, the LFNST was enabled, as long as the
primary transform pair is a combination of DCT2 and DST7,
the DST7 is changed to DCT2, and then does the additional
processing.
II. C URRENT COMBINATION OF IMPLICIT MTS AND LFNST
The combination of implicit MTS and LFNST in the
VTM5.0 reference software of the next generation video
coding standard is shown in Fig. 1. When implicit MTS is
enabled, initialize the type of horizontal and vertical
149