An Improvement of Unsymmetrical-Cross Multi-Hexagon-grid Search algorithm

Abs the (UM Des new very Key mot B a te prev bes mea squ bec O exh find com app algo ada mot qua be rapi I ada bes diff it h Thi mot mot the T be algo (FD UM com mot

An

M

stract— In this Unsymme MHexagonS) a scent Search (F w algorithm ca

y minor bit rat

ywords

—

Block matchin echnique used

vious frame f t matching b asurement, su uare error (M cause of its sim Obviously, the haustively sea

d the best mat mputational op plication. For y orithms ([1]-[ apt different

tion estimatio ality and the b

trapped in lo id motion scen In H.264 Re apted as the m

t matching b ferent kind of has the early te is algorithm

tion video seq tion vectors,

same as the fu Though the pe

improved a li orithm called DGDS) ([11]) MHexagonS.

mpressed vide tion scenes.

Impro

Multi-H

Vu T

Softwar

s paper, we pr etrical-cross

lgorithm and a FDGDS). The an reduce the m

te increase and

eo compressio n, video coding

I. INTR

ng motion est d to find the b from the im a block is foun uch as su m o SE). Howeve mplicity and e e simplest met arches all the

tching candid peration and years, research [9]) to tackle

search point on time with bit rate. These cal minimum nes.

eference Soft main search al lock in searc f patterns to fi ermination str works very quences. It ru

and good rate full search algo erformance of

ittle bit more d Fast Direct ). This new However, it eo, especially

oveme

Hexag

re Engineering

resent an new Multi-Hexag and the Fast D experiment re motion estimat d almost the sa

n, block matc g, UMHexagon

RODUCTION

timation in vid est matching age block in c

nd by using of absolute ( er, the SAD i

fficiency. thod is full se

points in the date. However is not feasible hers have prop the problem o patterns, try hout affecting algorithms, h m point, especi tware, UMHe lgorithm that

h window. T ind the best m rategy to redu well with bo uns quite fast

e-distortion p orithm. f UMHexagon

. Lai-Man Po ional Gradien algorithm run t increases y video sequ

ent of

gon-gr

g Department, Ho Chi Min

algorithm bas gon-grid S Directional Gra

esults show tha tion time with ame image qua

ching, improve nS.

deo compress image block current frame certain simi SAD) or me is commonly earch algorithm

search windo r, it requires l

e in most of posed many h of full search, ying to reduc g much the however, can e

ially in video exagonS ([10

is used to fin This algorithm matching block uce the search oth slow and and gives acc erformance, n nS is good, it o introduces a

nt Descent S ns faster tha the bit rat uences with

Unsym

rid Sea

, University of nh City, Vietna

sed on Search adient at this just a ality. ement, ion is in the e. The ilarity an of used m that ow to ots of video hybrid , each ce the video easily o with 0]) is nd the m uses k, and time. d fast curate nearly could a new Search an the te of rapid

By apply UMH our i abou rate-d Th UMH descr UMH prese A. U Th diffe recta exten demo prese searc the e Fig. termi Th algor predi refer termi

mmetr

arch a

)_{, Phan Cong}

f Information am

y integrating ying new HexagonS alg

improvement ut 10% to 30%

distortion perf he paper is HexagonS and

ribe two HexagonS alg ented to demo II. UMH UMHexagonS his algorithm rent search p angular full s nded hexagon onstration of ented with th ch point is (0 early terminat 2 shows each ination is used

Fig.1. Search

here are fou rithm for pred iction, corresp

ence picture p ination are de

rical-C

algorith m

g Canh(3)

Technology

the FDGDS search patte gorithm. The can speed u % without m formance.

organized a d the FDGDS

improvement gorithm. Final onstrate the eff

EXAGONS AND

Algorithm m consists of

pattern: unsym search, uneve n based search f search pro he search ran , 0). The U M tion strategy t h step in UMH

d.

h process of UMH

ur methods diction. They a

ponding bloc prediction. A scribed in ([10

Cross

h

S into the U ern, we can

experiment r up the motion much loss o f v s follow: In S are pr esente s will be lly, the exper ficient of prop

D FDGDS ALG

many steps, mmetrical-cro en multi-hexa h, diamond se ocedure in U nge equals 16 MHexagonS al

to reduce the HexagonS and

HexagonS algorit

employed in are median pr ck prediction ll these metho 0]).

UMHexagonS n improve results show n estimation t video quality n section II,

ed. Next, we employed riment results posed algorith

GORITHM

, each step u oss search, s m agon-grid sea earch. In Fig. UMHexagonS 6 and the in lgorithm also e estimation ti d where the e

thm, W = 16

n UMHexag rediction, up l and neighbo ods and the e

and the that time and the will in s are hm.

B. T (OT dire In F algo hor vert dire

SAD is SAD is small

Fig. 2

FDGDS Algo The FDGDS a TS). In OTS, ection until th Fig.3, an exam orithm ([1]) rizontal direc

tical direction ection.

S

Initial search 

Early te

Early te Unsymmetri

Small rectan

Uneven multi‐h

Extended hexag

Diamond it

Fi SAD is

SAD is s small

l

2 Flow chart of U

orithm algorithm is b

, the search e minimum po mple of the f

is shown. Fi tion, and t he n from the min

Start

point prediction

ermination

ermination ical‐cross search

gular full search

exagon‐grid search

gon iterated searc

terated search

inish s very big

s very big SA

UMHexagonS alg

ased on the on is continued oint along tha first OTS-bas irst, the OTS en the OTS nimum point f h

h SAD is big

AD is big

gorithm

ne-at-a-time s in one part at direction is ed block mat S is pe rforme

is performe found on horiz

search ticular found. tching ed on ed on zontal

FD cente follo

Ste of t CUR Ste CUR

direc relati < T, and searc

Ste to St

Ste one round corre

Ste motio Th

Ac often motio demo Th the How espec

F

DGDS uses O er to s earch f

ws:

ep 1: Calcula the search RRENT_MIN

ep 2: For eac RRENT_MIN: (a) perfor (b) set t ction as a DIR (c) if DIR ive rate distor

update CURR repeat Step ches).

ep 3: If no po ep 5; Otherwi ep 4: DIREC is set as CUR d. Go to Step esponding pos

ep 5: The alg on vector poin he RDR for a

ccording to ([ n used becau on estimatio onstrates the e he FDGDS al UMHexagonS wever, the bit

cially the vide

Fig. 3 Example o

OTS strategy for the minim ate the block d

window cen ch of the eigh :

rm point-to-po the minimum RECTIONAL_ RECTIONAL rtion (RDR) fo RENT_MIN w 2 (i.e., ski oint with DIRE

ise go to Step CTIONAL_MI RRENT_MIN p 2 with u pd sition.

gorithm is com nting to the po

certain directi

[11]), the thre use it will ke on time and example of FD lgorithm is sh

S algorithm t rate o f vid eo sequences w

of OTS algorithm

on eight direc mum point. Th

distortion mea nter and set ht directions oint directiona m BDM foun

_MIN; L_MIN is fou

or the current with this DIRE

ip the rema ECTIONAL_

4.

INs are comp N. This is th e dated CURRE mpleted. Retu osition with C ion is calculat

_ _ eshold value T

eep the balan d the video DGDS algorith hown to be m

in motion deo sequence with rapid mo

m

ctions around he algorithm i asurement (BD

t the value of the point w al OTS;

d in the cur und, calculate direction. If R ECTIONAL_M aining directi _MIN is found pared. The l ow e end of a se ENT_MIN and

urn with the f CURRENT_M

ted as follow:

T equals to 0 nce between o quality. F

hm.

more efficient estimation ti es also incre otion scenes.

d the is as DM) as with

rrent e the RDR MIN onal d, go west arch d its final MIN.

T UM usin sear patt new com A I sear min usin to q com term sear adv FDG pro

S and

S CU

dire earl per RD DIR rem

F

The proposed MHexagonS a

ng unsymmet rch, the modi tern in uneven w octagon/dec mplexity. Thes A. A Modified

In UMHexag rch and small nimum point

ng these two s quickly find t mputational tim mination has

rch and sma vantage of the

GDS. The ne posed algorith Step 1: Calcu d set the value Step 2: For ea URRENT_MIN (a) perf (b) set ection as a DIR

(c) if D ly termination formed, calcu DR < T,

RECTIONAL maining directi

Fig.4.. Example o

III.PROPOSE

d algorithm algorithm in t

trical-cross se ified FDGDS n multi-hexag

cagon pattern se two aspects d FDGDS Algo gonS algorit l rectangular from the ini search method the minimum me. Moreover been used in ll rectangular e early termin ew FDGDS hm is as follow ulate the BDM as CURREN ach of the eig N:

form point-to-the minimum RECTIONAL DIRECTIONA n strategy. If ulate the RD update C L_MIN and r ional searches

of FDGDS algorit

ED ALGORITHM

is different two aspects. earch and sma

is used. Seco on-grid search n to reduce s are described orithm thm, the un

full search ar tial point. Bu ds, the FDGD

point on this r, in the origin n between un r full search. nation, it is i with early te ws:

M of the sear T_MIN ght directions -point directio m BDM fou L_MIN; AL_MIN is f o

early termina DR for the c u CURRENT_M

repeat Step s).

thm

M

from the or Firstly, inste all rectangula ondly, the hex h is replaced w the computa d in the follow

nsymmetrical-re used to fin ut here, inste S algorithm is step to reduc nal algorithm, nsymmetrical-. In order to integrated int ermination us rch window c s of the point onal OTS; und in the cu ound, try usin

ation could n urrent directio MIN with

2 (i.e., ski p riginal

ad of ar full xagon with a ational wing.

-cross nd the ead of s used ce the early -cross o take to the sed in center t with

urrent ng the not be on. If this p the

Ste to St

Ste one round corre

Ste motio

B. Th the u step to loo Li H algor 16-po our s bit ra with

ep 3: If no po ep 5; otherwis ep 4: DIREC is set as CUR d. Go to Step esponding pos

ep 5: The alg on vector poin

Fig.5. H

Octagon and he second imp uneven multi

adopts the hex ok for the min Hong-ye introd rithm ([12]) th oint hexagon simulation sho

ate of compr rapid motion

oint with DIRE se go to Step 4 CTIONAL_MI RRENT_MIN p 2 with u pd sition.

gorithm is com nting to the po

Hexagon pattern i

d Decagon Pa provement ap i-hexagon-grid

xagon pattern nimum point i duces the new hat use the 8-p pattern to red ows that Li H

essed video l scenes.

Fig.6. Octag

ECTIONAL_ 4.

INs are comp N. This is th e dated CURRE mpleted. Retu osition with C

in UMHexagonS

attern pplied into the

d search step n with 16 poin in large area o w fast block m

point octagon duce the searc ong-ye’s meth larger than ex

gon pattern

_MIN is found pared. The l ow e end of a se ENT_MIN and

urn with the f CURRENT_M

algorithm

e algorithm is p. Currently, nts shown in F

of search wind motion estima pattern instea ch time. Howe hod can make xpected on v

d, go west arch d its final MIN.

I patt com The and Fig help num Wh app the pro

T Sof seq disa 28, New on

Instead of usi terns are u mputational tim

e first pattern d the second p .6. In our e x p cover most mber of search hen performed plied alternativ decagon patt posed algorith

The proposed ftware 9.6 w quence is IP

abled, one r ef RD Optimza ws, Coastguar the first 100

PER

UMHex Proposed a Proposed a speedup UMHex Bitrate in

UMHex Proposed a Proposed a speedup UMHex Bitrate in

Fig.7. Deca

ing one hexa used alternati me at this step is the octagon pattern is the d xperiment, usi points in the h points each d the search at

vely. First, the tern, and so o hm is shown i IV.EXPERIM

d algorithm is with the fol PPPIPPP…,

ference frame ation is turne d

rd, Foreman, frames. The r

RFORMANCE COM

P

xagonS 3

algorithm 3

algorithm p over xagon ncrease

P

xagonS 34

algorithm 34

algorithm p over xagon ncrease

agon pattern B

agon pattern, ively to fu p without muc n with 12 poin decagon with

ing these two search windo round is redu t this stage, tw e octagon patt on. The whole

in Fig.8

MENT RESULTS

s implemented llowing conf sub-pel moti e, Quantizatio d off. Video Stefan and C reference softw

MPARISON OF PROP Aki PSNR

(dB) (kbitBitra

9.83 413.2

9.83 413.1

9.94

- 0.024

Coastg PSNR

(dB) (kbitBitra

4.34 1880.

4.34 1880.

22.4

- 0.009

here two dif urther reduce ch bit rate incr nts shown in F 12 points sho o alternatively

ow, meanwhi uced from 16 t

wo patterns w tern is used, n e search proce

S

d in JM Refe figurations: ion estimatio on Parameter

sequences: A Container are t ware is runnin

POSED ALGORITH iyo

ate

t/s) ME time(s)

22 1.660

12 1.495

4%

4%

guard ate

t/s) ME time(s) .53 4.212 .37 3.267

44%

9%

fferent e the rease. Fig.5, wn in y will le the to 12. will be next is ess of

erence frame on is value Akiyo, tested ng on

Wind proce UMH PSNR bit ra

TABLE I HM WITH UMHEX

e PSNR

(dB) 36.58 36.57

C

e PSNR

(dB) 36.44 36.44

dows operatin essor. The p HexagonS alg R (dB), avera ate (kbit/s). Th

U SAD is sm SAD is sm

Fig.8. XAGONS ALGORIT

Foreman Bitrate

(kbit/s) ME(

872.28 3.20

875.62 2.46

23.28%

0.383%

Container Bitrate

(kbit/s) ME(

856.08 2.12

856.05 1.77

16.08%

- 0.004%

ng system w proposed algo gorithm in th age Motion E he experiment

Sta

Initial search po

Early term

Uneven multi‐octagon

Extended hexago

Diamond iter

Fini SAD is ve

mall mall

FDGDS with ear

. Search process o THM IN H.264 REF

E time

(s) PSNR(dB)

09 35.16

62 35.16

E time

(s) PSNR(dB)

20 38.26

79 38.26

with the Intel orithm is com hree different

stimation tim t results are sh

art

oint prediction

mination

n/decagon‐grid search

n iterated search

rated search

ish ery big

SAD is 

SAD rly termination

of proposed algor FERENCE SOFTWA Stefan R

) (kbit/s) Bitrate 2555.98 2557.31

27.62%

0.052%

News R

) (kbit/s) Bitrate 701.00 701.05

10.60%

0.007%

l Core i5-241 mpared with aspects: aver me (s) and ave r hown in Table

h big

D is big

rithm ARE

ME time (s) 3.928 2.843

ME time (s) 2.047 1.830

It can be seen that the proposed algorithm runs faster than UMHexagonS in all v ideo sequences, especially in video sequences with rapid mot ion scenes such as Coa stguard or Stefan. The bitrate of new algorithm is not much greater than the original algorithm with less than 0.5% increase, and it has even smaller bitrate in some videos such as Akiyo, Coastguard and Container. The PSNR does not change in most of the video sequences, except the Foreman with 0.01dB lost.

V. CONCLUSIONS

In this pape r, a new algorithm based on th e FDGDS and UMHexagonS algorithm is proposed. It takes advantage of the fast running time from FDGDS and accuracy from UMHexagonS to help reduce the computational complexity while maintaining the video quality and bitrate. And two new search patterns, octagon and decagon, are also presented to further reduce the motion estimation time. This new algorithm is shown to run fa ster than original UMHexagonS without much of PSNR lost and bitrate increase. It especially works well with videos with complex motion contents.

REFERENCES

[1] R. Srinivasan and K. R. Rao, “Predictive coding based on efficient motion estimation,” IEEE Trans. Commun., vol. 33, no. 8, pp. 888-896,

Aug. 1985.

[2] T. Koga, K. Iinuma, A. Hirano, Y. I ijima, and T. Ishiguro, “Motion compensated interframe coding for video conferencing,” in Proc. Nat. Telecommun. Conf., New Orleans, LA, Dec. 1981, pp. G5.3.1-G5.3.5.

[3] R. Li, B. Zeng, and M. L. Lio, “A new three-step search algorithm for block motion estimation,” IEEE Trans. Circuits Syst. Video Technol.,

vol. 4, no. 4, pp. 438-443, Aug. 1994.

[4] L. M. Po and W. C. Ma, “A nove l four-step search algorithm for fast block motion estimation,” IEEE Trans. Circuits Syst. Video Technol.,

vol. 6, no. 3, pp. 313-317, Jun. 1996.

[5] L. K. Liu and E. Feig, “A blo ck-based gradient descent search algorithm for block motion estimation in video coding,” IEEE Trans. Circuits Syst. Video Technol., vol. 6, no. 4, pp. 419-422, Aug. 1996.

[6] J. Y. Tham, S. Ranganath, M. Ranganath, and A. A. Kassim, “A novel unrestricted center-biased diamond search algorithm for block motion estimation,” IEEE Trans. Circuits Syst. Video Technol., vol. 8, no. 4,

pp. 369-377, Aug. 1998.

[7] C. H. Cheung and L. M. Po, “A novel cross-diamond search algorithm for fast block motion estimation,” IEEE Trans. Circuits Syst. Video Technol., vol. 12, no. 12, pp. 1168-1177, Dec. 2002.

[8] C. Zhu, X. Lin, and L. P. Chau, “Hexagon-based search pattern for fast block motion estimation,” IEEE Trans. Circuits Syst. Video Technol.,

vol. 12, no. 5, pp. 349-355, May 2002.

[9] C. H. Cheung and L. M. Po, “Novel cross-diamond-hexagonal search algorithms for fast block motion estimation,” IEEE Trans. Circuits Syst. Video Technol., vol. 7, no. 1, pp. 16-22, Feb. 2005.

[10] Z. B. Chen, P. Zhou, and Y. He, “Fast integer pel and factional pel motion estimation for JVT”, in Proc. 6th Meeting: JVT-F017, Awaji,

Island, Japan, 2002.

[11] Lai-Man Po, Ka-Ho Ng, Kwok-Wai Cheung, Ka-Man Wong, Yusuf Md. Salah Uddin, and Chi-Wang Ting, “Novel Directional Gradient Descent Searches for Fast Block Motion Estimation”, IEEE Trans. Circuits Syst. Video Technol., vol. 19, no. 8, August 2009.

[12] Li Hong-ye, Liu Ming-jun and Zhang Zhi-qiang, “A New Fast Motion Estimation Algorithm Based on H.264,” International Conference on Multimedia Information Networking and Security, vol. 1, pp. 287-290,