A fast H.266/QTMT intra coding scheme based on predictions of learned models

ABSTRACT

H.266/VVC adopts a QuadTree-plus-MultiType tree (QTMT) coding-unit (CU) split structure to improve efficiency at the cost of high time complexity. Speeding up VVC coding while minimizing quality degradation is critical for practical applications. We propose predicting the coding depth and split type of an optimally coded 32 × 32 CU (CU_32 × 32) to perform only a subset of exhaustive rate-distortion optimization (RDO) operations: (1) To predict the depth of an optimally coded CU_{32 × 32,} we train a convolutional neural network (CNN_depth). CNN_depth outputs a label specifying a depth range subset by which the controller can execute EarlySkip or EarlyTerminate to reduce time complexity. (2) To predict the split type, we train random forest classifiers (RFC_type). The corresponding RDO operations can be omitted if the RFC_type classifies one CU_32 × 32 as not of a specified split type. Experiments show that CNN_depth and RFC_type work seamlessly, reducing execution time by up to 69% and 39.16% on average, with a 0.7% increase in BDBR compared to the default VTM-7.0. Additionally, the proposed method yields the highest balanced time reduction rate of 61.5%.

CO EDITOR-IN-CHIEF:

• Su, Shun-Feng

ASSOCIATE EDITOR:

• Su, Shun-Feng

KEYWORDS:

• Versatile video coding
• quadtree plus multitype tree coding (QTMT)
• convolutional neural network
• random forest classifier

Nomenclature

BTD/QTD	=	Binary-tree/quadtree split depth
CCS	=	Coding control system
CUn×n	=	One CU with size n×n
CUk	=	K-th Coding Unit (32 × 32)
CUlow/CUhigh	=	One CU from a low/high-resolution video
CNNdepth(CUk)	=	Depth prediction model for a cuk
classdepth(CUk)	=	The largest BTD in a cuk
${\mathit{c}}_{{\mathit{y}}_{\mathit{i}}}$	=	The center feature vector of the $y_i$-th class (Equation 4(4) ${\mathrm{\ell }}_{\mathit{opt}}=\mathit{argmax}{\left\{{\mathrm{\ell }}_i\right\}}_{\mathit{i}\in \left\{0,1,\dots ,7\right\}}=\mathrm{CN}{\mathrm{N}}_{\mathrm{depth}}\left(\mathrm{C}{\mathrm{U}}^{\mathrm{k}}\right)$(4) )
DT	=	Decision tree
ES/ET	=	Early stop/early termination mode
HTT/VTT	=	Horizontal/vertical ternary tree split
HBT/VBT	=	Horizontal/vertical binary tree split
K	=	K: dimension of a label vector (Equation 4(4) ${\mathrm{\ell }}_{\mathit{opt}}=\mathit{argmax}{\left\{{\mathrm{\ell }}_i\right\}}_{\mathit{i}\in \left\{0,1,\dots ,7\right\}}=\mathrm{CN}{\mathrm{N}}_{\mathrm{depth}}\left(\mathrm{C}{\mathrm{U}}^{\mathrm{k}}\right)$(4) )
ℓi	=	Label of classdepth
ℓopt	=	The predicted label by cnndepth (cuk)
m	=	Batch size in the deep learning process. (Equation 4(4) ${\mathrm{\ell }}_{\mathit{opt}}=\mathit{argmax}{\left\{{\mathrm{\ell }}_i\right\}}_{\mathit{i}\in \left\{0,1,\dots ,7\right\}}=\mathrm{CN}{\mathrm{N}}_{\mathrm{depth}}\left(\mathrm{C}{\mathrm{U}}^{\mathrm{k}}\right)$(4) )
mi	=	The i-th classifier output of rfcmode
NO	=	Not split mode
QP	=	Quantizer parameter
QTMT	=	Quadtree plus multitype tree
RDC	=	Rate-distortion cost
RDO	=	Rate-distortion optimization
RFCtype(CUk)	=	The split type classification model
rfcqt	=	Binary classifier to perform QT or not
$\mathit{rati}{o}_{\mathit{useful}}^{m_i}$	=	Ratio of helpful classification of type mi
UVG266	=	A VVC encoder based on the Kvazaar HEVC encoder
VTM	=	VVC Test Model (reference software)
VVenC	=	Fast VVC encoder implementation
${\mathrm{W}}_j,\mathit{W}$	=	${\mathrm{W}}_j\in {\mathrm{R}}^d$ the j-th column vector of the weight matrix $\mathit{W}\in R^{\mathit{d}\times \mathit{n}}$ in the last fully connected layer (Equation 4(4) ${\mathrm{\ell }}_{\mathit{opt}}=\mathit{argmax}{\left\{{\mathrm{\ell }}_i\right\}}_{\mathit{i}\in \left\{0,1,\dots ,7\right\}}=\mathrm{CN}{\mathrm{N}}_{\mathrm{depth}}\left(\mathrm{C}{\mathrm{U}}^{\mathrm{k}}\right)$(4) )
${\mathit{x}}_{\mathit{i}}\in R^d$	=	The i-th input feature vector (Equation 4(4) ${\mathrm{\ell }}_{\mathit{opt}}=\mathit{argmax}{\left\{{\mathrm{\ell }}_i\right\}}_{\mathit{i}\in \left\{0,1,\dots ,7\right\}}=\mathrm{CN}{\mathrm{N}}_{\mathrm{depth}}\left(\mathrm{C}{\mathrm{U}}^{\mathrm{k}}\right)$(4) )
${\mathit{y}}_{\mathit{i}}$ and ${\mathit{y}}_{\mathit{i}}^{\mathit{L}\mathit{S}}$	=	One-hot and smoothed label vectors of $x_i$ (Equation 5(5) $\mathrm{L}={\mathrm{L}}_{\mathrm{S}}+\mathit{\lambda }{\mathrm{L}}_{\mathrm{C}}={\sum }_{\mathrm{i}=1}^{\mathrm{m}}log\frac{{\mathrm{e}}^{\left({\mathrm{W}}_{{\mathrm{y}}_{\mathrm{i}}}^{\mathrm{T}}{\mathrm{x}}_{\mathrm{i}}+{\mathrm{y}}_{\mathrm{i}}\right)}}{{\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{e}}^{\left({\mathrm{W}}_{\mathrm{j}}^{\mathrm{T}}+{\mathrm{b}}_{\mathrm{j}}\right)}}+\frac{\mathit{\lambda }}{2}{\sum }_{\mathrm{i}=1}^{\mathrm{m}}||{\mathrm{x}}_{\mathrm{i}}-{\mathrm{c}}_{{\mathrm{y}}_{\mathrm{i}}}|{|}_2^2,$(5) )
$y_{\mathit{k}\left(\mathit{i}\right)}\in y_i$	=	The k(i) is the index of correct label of $x_i$ And $y_{\mathit{k}\left(\mathit{i}\right)}$ = 1 (Equation 5(5) $\mathrm{L}={\mathrm{L}}_{\mathrm{S}}+\mathit{\lambda }{\mathrm{L}}_{\mathrm{C}}={\sum }_{\mathrm{i}=1}^{\mathrm{m}}log\frac{{\mathrm{e}}^{\left({\mathrm{W}}_{{\mathrm{y}}_{\mathrm{i}}}^{\mathrm{T}}{\mathrm{x}}_{\mathrm{i}}+{\mathrm{y}}_{\mathrm{i}}\right)}}{{\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{e}}^{\left({\mathrm{W}}_{\mathrm{j}}^{\mathrm{T}}+{\mathrm{b}}_{\mathrm{j}}\right)}}+\frac{\mathit{\lambda }}{2}{\sum }_{\mathrm{i}=1}^{\mathrm{m}}||{\mathrm{x}}_{\mathrm{i}}-{\mathrm{c}}_{{\mathrm{y}}_{\mathrm{i}}}|{|}_2^2,$(5) )
∆T and$\mathit{T}=\frac{\mathrm{\Delta }\mathit{T}}{\mathit{BDBR}}$	=	Time reduction rate and its balanced one

Acknowledgment

The authors express their gratitude to the reviewers for their insightful comments and suggestions, which have greatly improved this manuscript. Special thanks to Mr. Chong-How Fan for his assistance with editing and formatting during the revision process.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

We also acknowledge the support from the Taiwan NSTC project [NSTC 112-2221-E-011-113] for funding our research.