Professional Documents
Culture Documents
Lossless Image Compression Using Ideal Crosspoint Regions and The Integer Discrete Cosine Transform
Lossless Image Compression Using Ideal Crosspoint Regions and The Integer Discrete Cosine Transform
+
i 1
1 i
0 " 2
1
2
1
,
c i 0
.
'10)
The !unction ]
%
i *e!ine* a the greatet integer le than
or e;ual to %, +here % i a real num(er, 23 A 0 an* 2c - ", " i
the cale !actor that i mentione* (elo+.
Here i the co*e generation $roce*ure to e4act arithmetic
mani$ulation o! integer !or avoi*ing any inaccuracy +hich
might (e intro*uce* into the algorithm (y in*icriminate ue
o! !loating#$oint calculation. 8o at $reent, +e have to acce$t
an a$$ro4imation !or com$uting 2i (y uing $ro(a(ilitie 0i
that are real num(er +hich are a$$ro4imate*. &n thi
$roce*ure the cale !actor " that e!!ectively convert a real
$ro(a(ility to a !re;uency rate $er " ource ym(ol. Thi
cale !actor " can (e choen to avoi* the roun*ing e!!ect.
&n 910:, a!ter the $roce o! counting the ;uantity o!
a$$earance o! ource ym(ol +e have a et N A BN1, N2, C ,
NcE, +ith 0 O Ni O ", an* +ith ome im$le o$eration +e can
ee that the e;uation '10) (ecome
2i-
i 1
1
N
1
, c i 0 . '11)
8o the $roce*ure to com$ute 2i i not an a$$ro4imate
calculation an* n i the num(er o! ource ym(ol in thi cae,
(ut in ome cae i! +e alrea*y ha* the $ro(a(ilitie 0i, +e
coul* ue '10), not '11), to calculate 2i. An* i! +e ue '10), +e
have to choice the cale !actor " a$$ro$riate to the value o!
the $ro(a(ilitie 0i .
The (aic $rinci$le o! ource co*ing cheme i that (oth
u$$er an* lo+er limit o! the interval corre$on*ing to the
tring N are re$reente* (y three !unction 4(5, 6(5 an* L(5.
L(5 i in the com$onent o! e4$onent #L(5-w (y +hich the
(ae d i raie* to give a cale !actor. 4(5 an* 6(5 re$reent
the $oition an* +i*th, re$ectively, o! the interval, they are
integer, an* converte* into !raction (y the cale !actor
*
#L(5-w
.
Kith the initial con*ition
4( - 0, 6( - d
w
, L( - 0 , '12)
4, 6 an* L are *e!ine* recurively (y the !ollo+ing
e;uation '12), '1<) an* '1=)5
( )
s
i i
d " 2 6 4 4
,
_
1
]
1
+ +
2
1
J ) ' ) ' ) a '
1
,
'12)
( ) ( )
s
i i i
d " 2 6 " 2 6 6
,
_
1
]
1
+
1
]
1
+
2
1
J ) '
2
1
J ) ' ) a '
1
, '1<)
L'5ai A L'5 $ s, '1=)
+here 5ai i an e4ten*e* e;uence the interval limit o! +hich
are *erive* (y a ma$$ing o! the ne+ ym(ol ai onto the
$reviou interval (ae* on the cumulative !re;uency ta(le, s i
the integer !or +hich
d
w
O 6 '5ai ) O d
w$ 1
'1G)
The $arameter w i ue* to $reent the +i*th o! an interval
(y 6 '5), it e!!ectively *etermine the num(er o! d#ary *igit o!
$reciion (ecaue in the !loating#$oint term the caling (y d
s
in '1<) to ati!y '1G) i e;uivalent to a normali0ation o! 6
'5ai) to 'wD1) d#ary *igit.
,ei*e that, Jone 9I: $reente* main !eature, $ro$ertie,
enco*ing an* *eco*ing algorithm that +e ue* to *evelo$ our
cheme, (ut intea* o! uing the $ro(a(ilitie to com$ute the
cumulative !re;uencie +e ue the ;uantitie o! a$$earance o!
ym(ol, o +e *onPt carry out a$$ro4imate calculation !or
com$uting the cumulative !re;uencie 2i .
I). SC*E&E ICRIC 'OR LOSSLESS I&AE
CO&PRESSION
?igure 1. $reent the cheme &C%&C !or image com$reion
generally. 6ach te$ i num(ere* accor*ing to the e;uence o!
the cheme, +e have G te$ !rom 1 to G. The !irt te$, 8te$ 1
'&*eal cro $oint region) loo1 !or region o! cro $oint
+here +e can o$timi0e the $ro(a(ility o! *ata (it. -rey
value o! *ata $oint in thee region ati!y '1) !or &C%. 8te$
2 'Cro $oint ma$) +ill eta(lih the ma$ o! i*eal cro $oint
region. Notice that +ith each central value 2
n
+e have many
*i!!erent i*eal cro $oint region *o'n, p) on (it $lane 'n-p),
in thoe region +e enco*e *ata (it. Thi te$ evaluate an*
ignore mall i*eal cro $oint region containing 1, 2 or 2
*ata $oint *e$en*ing on the num(er o! (it $lane 'n#p), thee
area *onPt a!!ect much the com$reion ratio o! $rocee o!
co*ing an* *eco*ing. The thir* te$, 8te$ 25 -ray co*ing,
ma1e -ray co*e tran!ormation 92:. 8te$ < ',it $lane
*ecom$oition) *ecom$oe image *ata into e$arate (it
$lane that are num(ere* !rom 0 to N#1, thee num(er *e$en*
on the igni!icance o! (it, +here N i the (it length o! $i4el
o! image. The ne4t te$, 8te$ =5 7$timi0ing $ro(a(ility,
calculate the $ro(a(ility o! *ata (it outi*e &C% an*
o$timi0e $ro(a(ilitie o! *ata (it in &C% (y the $ro$oition
2.2. Thi o$timi0ation i (ae* on the cro $oint ma$
(e!orehan* an* im$lemente* to 8te$ G to com$ute !re;uencie
o! *ata (it. The lat te$, 8te$ G5 Co*ing, ue ome algorithm
in the $roce o! entro$y co*ing, li1e arithmetic co*ing. &n the
e4$erimental reult at the ne4t ection, +e ue JoneP
algorithm 9I: mentione* in the ection 2 a(ove. The $roce o!
enco*ing *ata (it houl* (e carrie* out !rom the mot to the
leat igni!icant (it $lane (ecaue o! the ran*om o! *ata on
the le igni!icant (it $lane.
There!ore, thi cheme i a $roce o! entro$y co*ing5
mo*eling '!rom 8te$ 1 to 8te$ =), an* co*ing '8te$ G).
2
2ig"re 1. The scheme &'(&' with ideal cross point regions
for lossless image compression.
). CONCLUSION
The cheme &C%&C an* ome reult are $reente*. Thi
cheme +ith the *e!inition 2.1, the $ro$oition 2.2, an* the
cone;uence 2.2 i no+ a ne+ $art o! the theory o! cro $oint
region +hich are mentione* in 912:, an* 91<:. &t illutrate* the
ue o! the theory o! cro $oint region +ith o$timi0ing
$ro(a(ilitie o! *ata (it only in i*eal cro $oint region. &n
thi $a$er, +e have ue* i*eal cro $oint region on (it $lane
'n - p), thi i the *evelo$ment !rom 91:, 92: an* 912:, 91<:. Ke
may chooe the num(er o! (it $lane 'n - p) (e!ore enco*ing,
that may ma1e the com$reion ratio *ecreae *ue to (igger
an* (igger cro $oint ma$, (ut thi i (etter an* (etter !or
encry$tion. The *ata receive* !rom the cheme i reveri(le,
o thi cheme can (e trongly ue* in cry$togra$hic ytem
+ith 1ey *e$en*ing on n#or*er etimating, num(er o! (it
$lane, num(er o! i*eal cro $oint region, thi $ro(lem +ill
(e mentione* in another $a$er. -enerally, the cheme &C%&C
i a $roce o! entro$y co*ing, it inclu*e t+o $art5 mo*eling
'8te$ 1#=) an* co*ing '8te$ G). The theory o! cro $oint
region can (e ue* in the !irt $art in or*er to re*uce
inter$i4el re*un*ancy, the econ* $art ue ome algorithm,
li1e arithmetic co*ing or JoneP metho* to re*uce co*ing
re*un*ancy.
The (aic conce$t have (een intro*uce* in 91:, 92: +hich are
no+ the (ae !or the theory o! o$timi0ing $ro(a(ilitie o! cro
$oint +ith the ne+ conce$t o! i*eal cro $oint region. A
meaning!ul im$rovement in com$reion ratio ha (een
o(taine*, com$are* to the other authorP metho*. ?rom thee
concrete (ae, the $ro(lem o! im$roving the com$reion
ratio o! image $roceing an* tranmiion can !urtherly (e
*evelo$e* in the !uture.
RE'ERENCES
91: Dang Thanh Tin, "u Dinh Thanh, Nguyen Qim 8ach,
8eiha1u Higuchi, R7$timi0ing $ro(a(ility o! *ata (it (ae*
on the theory o! cro $oint regionS, M&& .ournal, vol.22,
No. 2, May 200=.
92: Dang Thanh Tin, "u Dinh Thanh, Nguyen Qim 8ach,
8eiha1u Higuchi, R6!!ect o! -ray co*ing on the tate an*
the *itri(ution o! *ata (it in cro $oint region,S in 3roc.
2r* &nt. Con!. Com$utational 6lectromagnetic an*
A$$lication, &666 catalog No. 0<6T>H=, ,e.ing, China,
Nov 200<, $$. ==2F==G.
92: C. J. Lu, 8. C. Tai, RA note on iterating an #ary -ray
co*e,S 8&AM J. Dicrete Math, vol. 1<, No. 2, $$. 22H#22I,
2001.
9<: Nguyen Qim 8ach, RDigital image an* vi*eo
$roceingS, 8cience an* Technology 3u(liher, "ietnam,
1IIH, $$. 1H0#1H<.
9=: M.C. 6r, R7n generating the N#ary re!lecte* -ray co*e,S
&666 Tran. Com$uter, vol. 22, $$. H2I#H<1, 1I><.
9G: K. -il(ert, RA cu(e# !illing Hil(ert curve,S Math. &ntell. ,
vol. G, $$. H>#>1, 1I><.
9H: Matthe+ ,. 8;uire, R-ray co*e an* e!!icient generation
o! com(inatorial tructure,S Doctor thei, De$t. Com$uter
cience, North Carolina 8tate Univerity, 2001.
9>: %a!ael C. -on0ale0, %ichar* 6. Koo*, RDigital
image $roceingS, 3rentice Hall, 2002, $$ <2># <<0.
9I: Jone C. ,, R6!!icient co*ing ytem !or long ource
e;uence,S &666 Tran. &n!orm. Theory, vol. 2H, No. 2, $$.
2>0#2I1, 1I>1.
910: Dang Thanh Tin, "u Dinh Thanh, Nguyen Qim 8ach, RAn
im$rovement o! JoneU metho* o! lole *ata
com$reion,S "ietnam $ot an* telecommunication
.ournal, No. H, $$. G=#GI, May 2002.
911: ?e*eral in!ormation $roceing tar*ar* $u(lication 1IH,
RAnnouncing the A*vance* 6ncry$tion 8tar*ar*S, National
&ntitute o! 8tar*ar* an* Technology 'N&8T), Nov 2001.
912: Tin1u Acharya, 3ing#ing Tai, RJ36-2000 8tar*ar* !or
image com$reionS, Kiley, 200=.
912: Dang Thanh Tin, 8eiha1u Higuchi, RA lole co*ing
cheme !or image,S Me*ical &maging an* &n!ormation
cience .ournal, Ja$an, vol.1>, No. 2, $$. H=#>2, May 2001.
91<: Dang Thanh Tin, "u Dinh Thanh, Nguyen Qim 8ach,
8eiha1u Higuchi, RLole image com$reion +ith the
theory o! cro $oint region,S $reente* at <th &666 &nt.
Con!. Com$uter 8cience, %eearch, &nnovation an* "iion
!or the ?uture, Collection &n!ormati;ue &88N5 1G21#0>G=,
Hor 8erie &88N5 1G21#00G=, Hochiminh city, "ietnam, ?e(
200G, $$. 1I#2=.
<
3A%T 25 &NT6-6% D&8C%6T6 C78&N6 T%AN8?7%M
I. INTRODUCTION
The *icrete coine tran!orm 'DCT) ue the conce$t o! the
tran!orm co*ing that ha (et $er!ormance in image
com$reion an* !iltering !or image. The (loc1 i0e o! mot
a$$ro$riate are > or 1G !or the tran!orm co*ing o! the
image *ata. There!ore, im$lementation o! the or*er#> an*
#1G DCT ha !at com$uting time an* cot#e!!ectivene
!or reali0ation o! a tran!orm co*ing. Ho+ever, the
com$onent o! the (ai vector o! the DCT e4it
irrational num(er then cannot (e re*uce* to integer (y
im$le caling. There!ore, it i har* to im$lement an* uing
!loating $oint arithmetic i com$le4 an* e4$enive, o integer
*icrete coine tran!orm '&ntDCT) are $ro$oe* to
im$lement the DCT (y uing im$le integer arithmetic. 7n
the (ai o! theory o! *ya*ic ymmetry, tran!orm the or*er#
> an* or*er#1G coine tran!orm into a !amily o! integer
*icrete coine tran!orm '&ntDCT). &! the (etter
$er!ormance i *eman*e*, the large magnitu*e can (e ue*
!or the integer tran!orm.
&n thi econ* $art, +e *icu (rie!ly DCT tran!ormation
an* then ho+ to im$lement an* a$$ly &nteger DCT in lole
image com$reion.
II. DISCRETE COSINE TRANS'OR& +DCT,
The DCT ha long (een ue* a a metho* !or image co*ing
an* ha no+ (ecome the tan*ar* !or vi*eo co*ing . &t energy
com$action ca$a(ility ma1e it i*eal !or e!Vcient
re$reentation o! image. -iven a ;uare image (loc1 ? o! i0e
m (y m, an m (y m matri4 C i *eVne* (y the e;uation5
Ci.A
1
m
iA0 .A0..m#1
2 '2 1)
o
2
i 1
c i
n m
+
A1..m#1 .A0..m#1
Thu the DCT o! ? i *eVne* a !AC?C
T
'1)
The DCT i a unitary tran!orm,meaning that the inverion
an (e accom$lihe* (y ?AC
T
!C '2)
Un!ortunately, the DCT coe!Vcient, i.e., the entrie in !, are
evaluate* to inVnite $reciion. &n tra*itional co*ing metho*
(ae* on the DCT, all com$reion an* all loe are
*etermine* (y ;uanti0ation o! the DCT coe!Vcient. 6ven !or
lole image com$reion, thi $ro(lem cannot (e avoi*e*,
(ecaue toring the coe!Vcient to their !ull $reciion '+hich i
*etermine* (y the machine one i uing) +oul* not yiel* any
com$reion. Khat i $ro$oe* i to evaluate all entrie o! the
DCT matri4 out to only , *igit $at the *ecimal $oint. Thi
mean that the DCT coe!Vcient +ill have $reciion out to 2,
*igit $at the *ecimal $oint. A ma.or cone;uence o! thi
action i that the reulting DCT matri4 i no longer unitary,
an* the invere o! the DCT matri4 an* it tran$oe mut (e
evaluate* e4$licitly, i.e.,
?AC
#1
!'C
T
)
#1
'2)
2ig"re #. /ighest to lowest energy coef7cients.
7! coure, one mut chooe , uch that thee invere e4it.
?or intance, in the cae +here , i 1, i.e., the entrie o! C are
evaluate* to only one $lace (eyon* the *ecimal $oint, the
invere matri4 *oe not e4it in the > (y > $i4el (loc1 cae.
8ince the entire o$eration involve matri4 multi$lication, a
total o! 2m
2
multi$lication an* 2m
2
'm#1) a**ition are
re;uire* to evaluate all the entrie o! ' in '2).
7nce the DCT coe!Vcient have (een com$ute*, +e retain
only + high energy coe!Vcient to (e ue* !or the calculation
o! an a$$ro4imation to the original *ata matri4 ?. 7ne nee* to
chooe + uch that a *eire* amount o! energy com$action i
o(taine*. The high#energy coe!Vcient in general +ill a$$ear
in the ame location ini*e !/ e.g. the three highet#energy
coe!Vcient al+ay a$$ear in !00 , !01 , an* !11 , the u$$er#le!t
corner o! !. &n ?ig. 2, the high to lo+ energy coe!Vcient are
canne* !or an > (y > DCT (loc1. The remaining DCT
coe!Vcient are aume* to (e 0ero. Then the invere#DCT i
calculate* '+ithout auming unitarity) an* a reulting matri4
?n reult. ?rom thi, an error rei*ual matri4 6 can (e
*eVne*5
6i. A ?i. # ?ni. '<)
,y retaining E an* the + ;uanti0e* DCT coe!Vcient,
$er!ect recontruction o! ? can (e achieve*. A!ter electing
the high energy DCT coe!Vcient, +e $er!orm linear $re*
&ntDCTion on the non0ero DCT coe!Vcient (y uing a im$le
*i!!erencing cheme. ,et+een neigh(oring (loc1 there e4it
ome correlation (et+een the corre$on*ing high energy
coe!Vcient. A!ter $eci!ying + o! thee high energy
coe!Vcient, each o! thee coe!Vcient can (e enco*e* a the
error rei*ual reulting !rom u(tracting the corre$on*ing
DCT coe!Vcient !rom a neigh(oring (loc1/ in thi $a$er, the
(loc1 imme*iately to the le!t i choen, e4ce$t !or any o! the
le!tmot (loc1 in the image, +hich ue the (loc1 imme*iately
a(ove. A a reult, the high energy coe!Vcient are
*ecorrelate* (et+een (loc1, an* the overall Vrt#or*er
entro$y o! the reulting *ata i *ecreae*. &n a**ition, the
entro$y o! the entrie o! 6 are o! lo+er entro$y than the
original *ata matri4/ there!ore, a uita(le entro$y co*er can (e
ue*, uch a an a*a$tive Hu!!man co*er or an arithmetic
co*er.Moreover, ince +e are tranmitting error rei*ual, the
choice o! the $arameter ,, +hich *etermine the $reciion o!
the tranmitte* DCT coe!Vcient, (ecome le crucial, a
=
increaing , +ill reult in little *ecreae o! 6 in many cae.
?or intance, in the > (y > cae, e4$erimentation ho+e* that
, A 2 +a a*e;uate to achieve ma4imal com$reion 'a tate*
(e!ore, !or , A 1 the DCT matri4 invere *o not e4it in the >
(y > cae). The minimum choice o! , i
cae#*e$en*ent.
+-, Application of DCT in Image Compression
III. INTEER DCT
8everal verion o! the integer DCT have (een *evelo$e*.
The o(.ective i to im$lement the &NTDCT +ith integer
a**ition an* integer multi$lication '(it a** an* hi!t) an*
alo im$lementthe invere &NTDCT imilarly. A an e4am$le
everal '> 4 >) &NTDCT are lite* (elo+
#. Integer DCT .ia Lifting
To contruct &nteger DCT, the DCT 1ernel i *ecom$oe* into
-iven rotation via KalhFHa*amar* tran!orm. The KalhF
Ha*amar* tran!orm 'KHT) re;uire only a**ition
'u(traction) a all o! it element are 1 an* #1. The >#$oint
KHT matri4 i given (y5
A (it#reveral matri4 9+: i one +hich rearrange the in$ut
e;uence into the (it#reveral or*er.
Let B '0), '1), '2), '2), '<), '=), 'G), 'H)E
T
4 % % % % % % % %
then the $ro*uct
9 : B '0), '<), '2), 'G), '1), '=), '2), 'H)E
T
+ 4 % % % % % % % %
ha it entrie in (it#revere* or*er.
The DCT o! in$ut e;uence 4 i T A 9C:4, +here 9C: i the
DCT 1ernel. No+ let u normali0e 9H+: uch that
W
9 : 1J 9 :
w w
/ N / i an orthonormal matri4. Hence
1
W W
9 : 9 :
T
w w
/ /
]
an*
<<
co'HXJ1G) 0 0 #in'H J1G)
0 co'2XJ1G) #in'2XJ1G) 0
9 :
0 in'2XJ1G) co'2XJ1G) 0
in'HXJ1G) 0
8
0 co'HXJ1G)
0 #1 0 0 co'2XJ>) 0 #in'2XJ>) 0
0 0 0 1 0 co'2 J >) 0 #in'2
4 4
#1 0 0 0
0 0 #1 0
1
1
1
1
1
]
1
1
1
1
1
]
J >)
in'2 J >) 0 co'2 >J) 0
0 in'2 J >) 0 co'2 J >)
1
1
1
1
1
]
To o(tain &NTDCT, the li!ting !actori0ation o! -iven rotation
matri4 i ue*5
co 1 co 1
co #in 1 0 1 1
in in
in co in 1
0 1 0 1
1 1
1 1
1 1
1 1
1 1
] ]
] ]
T+o *itinct integer multi$ler an* are ue* to archieve
integer li!ting (y5
co 1
2 in
b
b
an* in
2
b
b
The !ollo+ing ta(le lit the multi$lier
b
an*
b
!or
(A1 # > an* variou rotation angle5
G
Thi !igure ho+ ho+ a -iven rotation i integer li!te*.
$. Integer DCT /0 the Principle of D0adic S0mmetr0
Another integer DCT can (e *evelo$e* (y re$lacing
tran!orm 1ernel com$onent o! the or*er#eight DCT (y a ne+
et o! num(er. ?or the or*er#eight DCT, many tran!orm o!
thi 1in* can (e o(taine*. ,oun*ary con*ition are a$$lie* to
enure that the ne+ tran!orm contain only integer, +hile
they till a$$ro4imate the DCT. A the magnitu*e o! thee
integer can (e very mall, the tran!orm can (e im$le to
im$lement.
Let the matri4 9T(: re$reent the 1ernel o! the eight#$oint
DCT. The !ollo+ing *ecri(e the te$ to convert the or*er#
eight DCT into integer DCT 1ernel.
&! the (ai ce4tor Jm 'mA0, 1,..., H) o! 9T%: are cale* (y
9Q:, all the (ai vactor can (e e4$ree* a the varia(le o! a,
(, c, *, e, ! an* g a !ollo+5
+here 9Q: A *iag '10, 11, ..., 1H) an* YY1mJmYY
2
A 1 'm A 0, 1, ...,
H).
?rom the !ollo+ing ta(le, the only con*ition to ma1e the
tran!orm 9T%: (e orthogonal i5 a( A ac D (* D c*
A there are < varia(le, there are an in!inite num(er o!
olution. Thu an in!inite num(er o! ne+ orthongonal
tran!orm can (e generate*.
?rom thee num(er +e can et u$ (oun*ary con*ition a
!ollo+ e4$ecting that the ne+ tran!orm a$$ro4imate the
DCT an* ha a goo* $er!ormance a the DCT5
a b c d an* e f
?urthermore to eliminate truncation error *ue to o$eration
o! irrational num(er the !ollo+ing con*ition ha to (e
ati!ie*5 a, (, c, *, e, !
Tran!orm that ati!y thee con*ition are re!erre* to a >#
$oint integer DCT. 7nce the varia(le o! 9J: are !i4e*, +e can
get the caling !actor that ma1e the ne+ tran!orm
orthonormal. Ke can get5
0 <
2 2 2 2
1 2 = H
2 2
2 G
1J '2 2 )
1J 2' )
1J '2 )
9 9 g
9 9 9 9 a b c d
9 9 e f
+ + +
'