An Improved Detection Method for Z Q ti d Bl k H 264/AVC Zero Quantized Blocks on H.264/AVC

An Improved Detection Method for Z Q ti d Bl k H 264/AVC Zero Quantized Blocks on H.264/AVC

Bo-Jhih Chen and Shen-Chuan Tai

Presenter: Bo-Jhih Chen Date: 16 November 2010

Outline

Introduction li i

Preliminary

Proposed Method Proposed Method

Experimental Results

Conclusions

Introduction

Motivation

A 4 b 4 id l bl k ( i d )

A 4-by-4 residual block (motion-compensated error) There are many zero quantized DCT coefficients

after DCT/Q, especially in low bit-rate video coding.

Objective Objective

Early detect more zero-quantized coefficients before DCT/Q

DCT/Q.

Effective reduction of computational complexity.

Introduction

An 4x4 integer transform in H.264

Post-factor matrix

=

= ⊗

F AXA W PF

A matrix of the floating A matrix of the integer Post factor matrix

Quantization

A matrix of the floating

transform coefficient A matrix of the integer transform coefficient

( , ) ( , ) ( , , )

F u v

= W u v MF QP u v ⋅ + >> k qbits

If |F_Q(u,v)| < 1,

then F(u v) is considered as

ZQB (Zero Quantized Block):

A DCT block consists of 16 zero

Note1 Def.

then F(u,v) is considered as

a zero quantized DCT coefficient.

A DCT block consists of 16 zero quantized DCT coefficients.

Introduction

The sufficient critera for determining a zero q anti ed DCT coefficient

quantized DCT coefficient.

For W(u,v) Integer transform

( , ) 2

( )

qbits

k W u v

MF QP

< −

g f

( , , ) MF QP u v

For F(u,v) Floating transform

( , ) 5

F u v < 6 Qstep ( , )

6 Q p

where 0.625 2Qstep = × ^QP/ 6

Introduction

For example:

Recall _{4 0 1 2 1 6}

F_MAX

at Direct Current position Recall

=

= ⊗

F AXA W PF

-4 0.1 2 1.6

-3 -2.8 1.1 1.66

F

-2 -3 0 0

0 -3 0 1

AXA

0 1.16 0.08 -0.7

F

Floating

0 3 0 1

2 0 -2 -1

1 2 1 0

= W PF ⊗

I

-21 -30 7 15 14 -8 -10 -4

An input residual

block X ^Integer W⊗PF

W_MAX 14 -8 -10 -4

-3 10 1 -5

W_MAX

Preliminaryy

Energy conservation

2 2

T

( , ) ( , )

x y u v

f x y = F u v

∑∑ ∑∑ ^{f T F}

DC’s Energy

1 1

1

0 0

(0,0) 1

( , )

DC

x y

E F f x y

N

− −

= =

= = ∑∑

ACs’ Energy

( )

∑∑ ^{F F N}

E

⁼ _∑∑ ^F ( ^{u v} )

^{− F} ( 0 0 )

⁼ ( ^N σ

)

E ( , ) ( 0 , 0 ) σ

Preliminaryy

The variance of the (u,v)

DCT coefficient

2 2

, ,

( , ) [

] [

]

F

u v

ARA

ARA

σ = σ

A is the DCT transform matrix.

R is the covariance matrix.

B th t l li it th

( ) 5 Q

By the central limit theorem,

( , )

F

u v 6 Qstep

γσ <

Note2 If γ = 3,

the probability of F(u,v) = 0 is about 99.73%.

Preliminaryy

Over 99% of DCT coefficients will be equal to ero if

to zero, if

5

⎛ Q ⎞

⎜ ⎟

2

2

5 6

[ ] [ ]

f T T

Qstep

ARA ARA

σ γ

⎛ ⎞

⎜ ⎟

⎝ ⎠

<

, ,

[ ARA ] [

ARA ]

γ

denotes the variance of input residual data.

2

σ

Proposed Method p

For AC components,

⎛ ⎞

2 2 2 2

( , ) (0,0)

ACs f

u v

E = ∑∑ F u v − F = N σ

2

2

5

6 Qstep

σ

⎛ ⎞

⎜ ⎟

⎝ ⎠

<

, ,

[ ] [ ]

f T T

u u v v

ARA ARA

σ ^< γ

22

2

5

6 ( )

N Qstep

E TH u v

⎛ ⎞

⎜ ⎟

⎝ ⎠

<

, ,

( , )

[ ] [ ]

ACs T T ACs

u u v v

E TH u v

ARA ARA

γ ^{⎝ ⎠}

< =

Proposed Method p

v

TH

is a symmetric matrix.

TH (u v) = TH (v u)

(0,0) (0,1) (0,2) (0,3) (1,0) (1,1) (1,2) (1,3)

u

TH

(u,v) = TH

(v,u)

Wang et al. 2007

Occurrence Probability (OP)

Loc(0), Loc(1), Loc(2), Loc(3)

Based on the analysis, OP is over 80

% on 3×3. Hence, the threshold of DC energy can be approximately DC energy can be approximately TH(0,2)=TH(2,0) instead of TH(u,v).

Proposed Method p

Our method for ZQB detection

1 1

1 5

(0,0)

( , )

E

= F = ∑∑

f x y < Qstep

2

2

⎛ 5 ⎞

0 0

( , ) ( , )

DC

6

x y

f y Q p

N ∑∑

DC part

ZQB ?

2

2

5 (0, 2) 6

[ ] [ ]

ACs ACs T T

N Qstep

E TH

ARA ARA

γ

⎛ ⎞

⎜ ⎟

⎝ ⎠

< =

0,0 2,2

[ ARA ] [ ARA ]

ACs part

γ

Experimental Results p

Performance evaluations

.

P P

P N

Δ = −

(%) 100%

z

DR N

= N ×

… Detection Rate N: # of the detected ZQBs N_z: # of the actual ZQBs

( ) ( )

(%)

100%

z n

N N N N

DQ N N

+ − +

= ×

+

z n Q y

N_n: # of the actual non-ZQBs N_z: # of the actual ZQBs

N_m: # of the miss-detected ZQBs N_f: # of the false-detected ZQBs

z Q _f Q

Experimental Results p

PSNR / Detection rate, Akiyo

Experimental Results p

PSNR / Detection rate, Foreman

Experimental Results p

Detection Quality

Akiyo Foreman

Experimental Results p

Computation saving rate

.

(%) (1 ) 100%

Org

CSR OP

= − OP ×

g

Conclusions

An insignificant PSNR degradation

O 0 08 dB d f l i h

On average, 0.08 dB drop of our algorithm.

Higher detection ratio g

On average, 90 % of detection ratio.

Especially encoding at lower QPs (high bit rate Especially encoding at lower QPs (high bit-rate coding)

C i l i

Computational savings

Up to 54.6 % of DCT/Q/IQ/IDCT computations

can be reduced by using our method.

Thank you for your listening. y y g