Adjustable Partial Distortion Search Algorithm for Fast Block Motion

Adjustable Partial Distortion
Search Algorithm for Fast Block
Motion Estimation
Chun-Ho Cheung and Lai-Man Po
Department of Electronic Engineering,
City University of Hong Kong
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY,
VOL. 13, NO. 1, JANUARY 2003
Outline


Introduction to partial distortion search
algorithm
Progressive partial distortion






Preliminaries
Fast BMA and Searching Speed Limitation
Formation of PPD
PPDS Algorithm
Adjustable partial distortion comparison
Experimental results
Partial distortion search algorithms
Alternating Subsampling Search
Algorithm (ASSA)
{Pixel-decimated (4 : 1)}
Normalized Partial Distortion Search
algorithm (NPDS)
{early rejection, halfway-stop}
Partial distortion search algorithms


Without limitation of checking points
Lack the flexible adjustability between the
prediction quality and searching speed

Adjustable Partial Distortion Search algorithm
(APDS)


Increase the searching speed of NPDS by introduction
of progressive partial distortions (PPD) at very early
stages.
Adjustable partial distortion comparison method with
enabling the quality/speed control.
Progressive Partial Distortion
- Preliminaries
The basic operations of computing SAE are absolute and addition operations,
and require about (3N 2 -1) operations per BDM.
Progressive Partial Distortion
- Fast BMA and Searching Speed Limitation

For one second of K-Hz I x J video sequence with
search windows ±W


3SS


K(I/N)(J/N)(3N2-1)(2W+1)2
Minimizing the searching points in (2W+1)2
ASSA and NPDS




Reduce the BDM’s (3N2-1)
ASSA => 4 times speedup
ASSA + subblock or subsampled motion field => 8 times
speedup
NPDS => 12-13 times speedup
Conventional partial distance search
algorithm
•Pixel-by pixel basis
•Obtains the optimal solution as in FS
•The basis for developing NPDS
S=4
1
T=4
9
3 13
11 5 15 7
4 14 2 10
16 8 12 6
P =16
NPDS


Saving of multiples of 16-pixel matchingoperations
It limits the maximum possible speedup ratios
to Num(block)/Num(d1) or 16 times theoretically.
Number of pixel in the candidate block

distortion
d first
PPDS are proposed to First
be partial
used
in the
few stages of NPDS for increasing the
rejection rate.
1
•Also, wider range of quality control
Formation of PPD
P = 16
partitions
•Theoretically, it is a combinational-nature problem to divided d1
into smaller partitions.
•Regularity of the PPD patterns favors both hardware and software
Implementations.
•Total G partial distortions
G=H+P-1
Proposed PPD patterns
H=16, G=31
6 14 8 16
H=8, G=23
H=4, G=19
H=2, G=17
3
7
4
8
2
4
2
4
1
2
1
2
10 2 12 4
5
1
6
2
3
1
3
1
2
1
2
1
7 15 5 13
4
8
3
7
2
4
2
4
1
2
1
2
11 3
6
2
5
1
3
1
3
1
2
1
2
1
9
1
(a)PPDS(v1)
Group of 1 pixel
(b)PPDS(v2)
(c)PPDS(v3)
(d)PPDS(v4)
Group of 2 pixels Group of 4 pixels Group of 8 pixels
4
6
4
6
3
5
3
5
1
5
3
5
2
3
2
3
5
2
5
3
4
1
4
2
5
4
5
4
3
1
3
1
4
6
4
6
3
5
3
5
3
5
2
5
2
3
2
3
5
3
5
1
5
2
4
1
5
4
5
4
3
1
3
1
(e)PPDS(v5)
(1,1,2,4,4)
(f)PPDS(v6)
(2,2,4,4,4)
(g)PPDS(v7)
(1,1,2,4,8)
(h)PPDS(v8)
(4,4,8)
H=6, G=21
H=5, G=20
H=5, G=20
H=3, G=18
An Example
•Total G partial distortions
G=H+P-1
For dg|1≤g≤4,
H=4, G=19
2
4
2
4
3
1
3
1
2
4
2
4
3
1
3
1
(c)PPDS(v3)
Group of 2 pixels
For dg|5≤g≤19, same as the pervious PPD Formulation.
PPDS Algorithm
•Normalized Distortion Comparison criteria (NDC):
Dg  DMIN , where Dg  Dg / n, DMIN  DMIN / N 2
N 2 Dg  f (n, k ) DMIN , for f (n, k )  n
•Adjustable function:
f (n, k )  (1  k )n  kN 2
f(n,k)=n, where n is the number of pixels cumulated in Dg
Experimental results
COMPUTATIONS AND PSNR (dB)
PERFORMANCE COMPARISON
COMPUTATIONS AND PSNR (dB)
PERFORMANCE COMPARISON
COMPUTATIONS AND PSNR (dB)
PERFORMANCE COMPARISON
Performance comparison of APDS(k) at
several quality factor k
PPDS(v3)
Performance comparison of APDS(k) at
several quality factor k
PPDS(v3)
Performance comparison of APDS(k) at
several quality factor k
PPDS(v3)
Average PSNR performance and speedup
ratios
SIF, “tennis”
Average distance from and probability of the true motion
vector per block against the quality factor k
SIF, “tennis”
Average PSNR performance and speedup
ratios
CCIR601, “tennis”
Average distance from and probability of the true motion
vector per block against the quality factor k
CCIR601, “tennis”
MSE performance
Operations
Distance per block
Probability per block
Operations Translation in RHS
Conclusions

NPDS + different PPD


Computational reduction up to 61.54 times with
less than 0.94-db degradation on PSNR
compared to FS’s.
The APDS is very suitable for a wide range of
video applications such as low-bit-rate video
conferencing and high-quality video coding.