View Single Post
Old 16th September 2019, 13:43   #155  |  Link
Jamaika
Registered User
 
Join Date: Jul 2015
Posts: 697
https://www.researchgate.net/publica...and_Complexity 10.2018
HEVC transform is based on a Discrete Cosine Transform (DCT-II), apart for Intra-coded 4x4 TUs for which a
Discrete Sine Transform (DST-VII) is used. In comparison, JEM relies on a variety of selectable core transforms from
the DCT and DST families (DCT-II, DCT-V, DCT-VII, DST-I and DST-VII). An Adaptive Multiple Transform (AMT)
method has been designed such that depending on the selected mode (intra or inter), or further on the intra prediction
direction, a sub-set of the core transforms is built and one transform from this sub-set is chosen based on RDO. In
the JEM, large block-size transforms, up to 128×128 in size, are enabled. High frequency transform coefficients are
zeroed out for the transform blocks with size larger than or equal to 64, so that only the lower-frequency coefficients
are maintained. In addition to AMT, a Mode-Dependent Non-Separable Secondary Transform (MDNSST) is applied
between the core transform and the quantization, with the motivation to reduce remaining dependencies after the
separable core transforms which only address horizontal and vertical signal dependencies. Finally a Signal
Dependent Transform (SDT) is competed to the AMT output. The SDT approximates the optimal Karhunen-Loéve
transform (KLT), which is a signal dependent transform, by estimating current signal to code (transform block) with
similar signals (i.e. reference patch) available at the decoder (already coded).
AV1 also supports multiple transforms: DCT, Asymmetric DST (ADST), flipped ADST, and Identity (IDTX). The
identity transform is equivalent to the transform skip mode of HEVC and JEM. The vertical and the horizontal
transforms can be selected independently from the set of four transforms, resulting in 16 transform combinations.


As to transform, the square transforms in HEVC are extended to non-square transforms for rectangular blocks resulted from binary and ternary tree splits. Besides, [VVC] supports multiple transform sets (MTS), including DCT-2, DST-7, and DCT-8 as well as the non-separable secondary transform. The transforms used in [VVC] can have different sizes with support for larger transform sizes. For DCT-2, the transform sizes range from 2x2 to 64x64, and for DST-7 and DCT-8, the transform sizes range from 4x4 to 32x32. In addition, [VVC] also support sub-block transform for both intra and inter coded blocks. For intra coded blocks, intra sub-partitioning (ISP) may be used to allow sub-block based intra prediction and transform. For inter blocks, sub-block transform may be used assuming that only a part of an inter-block has non-zero transform coefficients.

Increase max QP from 51 to 63. An enhanced rate distortion optimized quantization scheme called Dependent Scalar Quantization. CABAC coder from AVC, it has been enhanced and is now even faster.

Last edited by Jamaika; 16th September 2019 at 14:11.
Jamaika is offline   Reply With Quote