1994 I nt er nat i onal S y mp o s i u m on Speech, I mage Pr ocessi ng and Neur al Net wor ks , 13-16 Apr i l 1994, Hong Kong

Edge Det ect i on on Real Ti me Usi ng LOG Fi l t er

Shen Li qi n , Shen Di nggang and Qi Fei hu
Fi br e Opt i c Local Communi cat i on net wor ks and Coher ent Communi cat i on
St at e Key Labor at or y
Shanghai J i aoTong Uni ver si t y, 200052, Chi na
A s t r ac t:
I n t hi s paper a new t echni que i s pr oposed f or
LOG( Lapl aci an Of Gaussi an) f i l t er i ng based on t he
met hod us i ng l ook- up- t abl e. We t ake advant age of
t he ci r cul ar l y symmet r i c char act er i st i c of LOG
f i l t er , t hen a l ook- up- t abl e wi t h accept abl e si ze
i s es t abl i s hed. Thi s t r ansf or ms a l ar ge amount of
mul t i pl yi ng oper at i ons i nt o t he easy l ooki ng up a
t abl e oper at i ons . At t he same t i me , anot her mas k
whi ch can f or m an adapt i ve t hr eshol d i s added t o
t he LOG mas k so t hat t he deci si on whet her a edge
exi s t s or not f r om t he r esul t of convol ut i on can
be made eas i l y. As a r es ul t , LOG f i l t er i ng becomes
a met hod of edge det ect i on on r eal t i me.
1 Introduction
I mage s egment at i on i s of cr uci al i mpor t ance i n
an obj ect r ecogni z i ng syst em. Edge det ect i on i s
most f r equent l y us ed i n i mage segment at i on. Ther e
ar e t wo br oad cl as s es of edge det ect or s. 0ne cl ass
i ncl udes many ki nds of gr adi ent oper at or s such as
Lapl aci nn oper at or , Rober t oper at or and Sobel
oper at or et c. . These oper at or s ar e ver y sensi t i ve
t o noi se, so t he ef f ect of r esul t i s har dl y
s at i s f yi ng. LOG f i l t er i s devel oped as a best
f i l t er s i mul at i ng t o human bei ng s vi si on syst em.
The excel l ent edge i nf or mat i on can be obt ai ned
usi ng i t . But t he ef f ect of edge det ect i on i s
deci ded cr uci al l y by whet her t he sum of LOG mask
coef f i ci ent s, val ues of LOG f unct i on on di scr et e
poi nt s, i s z er o or not . I n or der t o ext r act edges
ef f ect i vel y, t he wi dt h of LOG mas k must be chosen
much l ar ger compar ed t o t hose of gener al gr adi ent
oper at or s So t he amount of oper at i ons of
comput i ng convol ut i on of t he or i gi nal i mage wi t h
a LOG mas k i s unendur abl e i n many appl i cat i ons.
especi al l y f or t he obj ect r ecogni zi ng syst em on
r eal t i me. Ther e i s a l ot of r esear ch wor k f or
f ast al gor i t hmof LOG f i l t er i ngI l 1 I21 I51 I 61. We I l
vi ew t hese met hods br i ef l y l at er . I n t hi s paper , a
new t echni que i s pr oposed f or LOG f i l t er i ng based
on t he met hod usi ng l ook- up- t abl e. We t ake
advant age of t he ci r cul ar l y symmet r i c
char act er i st i c of a LOG f i l t er , t hen a l ook- up
- t abl e wi t h accept abl e s i z e i s est abl i shed. Thi s
t r ansf or ms a l ar ge amount of mul t i pl yi ng
oper at i ons i nt o t he easy l ooki ng u p a t abl e
oper at i ons. At t he same t i me , anot her mas k whi ch
can f or man adapt i ve t hr eshol d i s added t o t he LOG
mas k so t hat t he deci si on whet her a edge exi st s or
not f r om t he r esul t of convol ut i on can be made
eas i l y. As a r esul t , LOG f i l t er i ng becomes a met hod
of edge det ect i on on r eal t i me.
Thi s paper i s ar r anged as f ol l ow: Sect i on 2 i s
t he anal ys i s of LOG f i l t er and t he amount of
oper at i ons of LOG convol ut i on , some pr evi ous
r esear ch f or f as t al gor i t hmof LOG f i l t er i ng i s
vi ewed , t oo. Then t he i mpl ement at i on of LOG
f i l t er i ng met hod usi ng t hr ee l ook- up- t abl es i s
pr esent ed i n Sect i on 3. Sect i on 4 i ncl udes some
exper i ment r es ul t s . The concl usi on i s dr awn i n
Sect i on 5.
2 LOG Filter and previous fast
algorithm of LOG convolution
LOG f i l t er i ng i ncl udes sequent i al pr ocessi ng of
s moot hi ng f or noi se r emove f i r s t l y and comput i ng
second der i vat i ve. Edges ar e i dent i f i ed as t he
l ocus of poi nt s wher e t he s econd der i vat i ve of t he
f i l t er ed i mage i s equal t o zr eo.
Fi g. I 2 - D pr of i l e of LOG mas k
0-7S03-1865-X/94/$3.00 0 1994 IEEE 37
ISSIPNN94
LOG f unct i on i n t wo di mens i ons i s gener al l y
expr essed as f ol l ow:
3G 3G
3 x 2 3 y a
( 2 - - ) exp( - - ) (1)
V" G( x , y , 6) = - t -
1 X 2 t f xat ya
=-
2 n 6 6 2 h a
I t ' s a ci r cul ar l y symmet r i c f unct i on shown as
Fi g, 1, The wi dt h of t he cent r al exci t at or y r egi on
of t he oper at or i s gi ven by w = 2 2 6 . H' hen
convol ut i on i s done on di gi t al i mages, t he si r e of
mas ks i s al ways l i mi t ed. Gener al l y, i f we chose
t he wi dt h of a mas k as f ol l ow:
t he cont i nuous cur ve of t he LOG f unct i on cover s
99. 7% of t he ar ea, gi ven an ar ea ver y cl ose t o
zer o f or t he V2 G. Thi s means t hat t he si ze of a
mask i s chos en based on t he chosen spat i al
const ant s 6.
Ber zi n[ 4J has done deep r esear ch on accur acy of
LOG edge det ect or and t he concl usi on was dr awn:
Ther e i s al i as i ng of geomet r i c l ocat i on usi ng a
LOG f i l t er as a edge det ect or . The l ar ger t he
spat i al cons t ant s i s , t he mor e ser i our al i asi ng
i s. Wi t h some cons t r ai nt s , t hi s al i asi ng i s not
mor e t han 6 , Fur t her mor e, t he maxi mum al i asi ng i s
bel ow 2 6 , As a r esul t , i t i s not sui t abl e t o
choose a l ar ge 6 i n or der t o obt ai n accur at e
l ocat i ons of edges.
On t he ot her hand, t he passband wi dt h of a LOG
f i l t er i s cons t r ai ned by 6 . A smal l er 6 i s
cor r espondent t o a wi de passband , as a r esul t ,
mor e det ai l ed i nf or mat i on of edges i s i ncl uded i n
t he f i l t er ed i mage. On t he cont r ar y, a l ar ge 6 i s
cor r espondent t o a nar r ow passband. S o t he
i nf or mat i on of edges i s much mor e br i ef .
Mu ~t i s c a ~i n g i mages r epr esent at i on can be obt ai ned
wi t h t he di f f er ent cons t ant s h . Gener al l y a
smal l er 6 i s chos en i n i mage segment at i on t o
assur e t hat al l usef ul i nf or mat i on i s i ncl uded i n
f i l t er ed i mage, But t oo smal l h makes t he
abi l i t y of noi s e i nhi bi t i on of LOG f i l t er t oo weak
Compr omi se s houl d be made on t he needs of
appl i cat i ons. I f we choos e
i-
N=3* 21 6 =8. 5 h
6 = 0.8 - 5
The si ze of mas k woul d be
N = 3 - 4 3
I t i s obvi ous t hat t he si ze of LOG masks
cor r espondi ng t o a l ar ge b i s much l ar ger t han
t hat of convent i onal gr adi ent oper at or s whi ch
r ar el y exceed
convol ut i ng of M*M i mage wi t hsuch N* N masks ar e
si ze of 5 x 5 pi xel s.
The t i mes of mul t i pl yi ng oper at i ons f or
( M- Nt l ) ' *NZ=WI N)
Fas t convol ut i on al gor i t hm f or LOG f i l t er must
be devel oped t o per f or m t he convol ut i ons i n a
r eas onabl e amount of t i me. I n paper [ Zl , t he aut hor
s ubs t i t ut ed DOG oper at or f or LOG oper at or .
1 x 2 t y 2
DOG( x, y , 6 ., 6 1) = ~ exp( - - )
3-7 6 e 2 6 2
I x2t y2
exp ( - - 1 12)
- -
i mhI 2 6 :
Wi t h a s ui t abl e pai r of h e and 6 I , t hi s DOG
f unct i on can mat ch LOG f unct i on expr essed as
f or mul a (1) accur at el y. Because Gaussi an f unct i on
i s a s eper abl e one, each of convol ut i on can be
r eal i z ed a s s ucces s i ve one di mensi onal r ow and
col umn convol ut i on, t hi s r educi ng t he t ot al
number of oper at i ons t o
(M-NI11 * 4N = 4MZN
The r educt i on of t h e t ot al number of
oper at i ons i s sl i ght . J . S. Chen[ l l et c. pr oposed t o
decompos e a LOG f i l t er wi t h spat i al const ant 6 t o
t wo s ucces s i ve f i l t er s, one i s Gaussi an f i l t er
wi t h 6 and t he ot her i s LOG f i l t er wi t h h 2 .
They f i r s t convol ved t he i mage wi t h a Gaussi an
f i l t er , t hen r educed t he r esol ut i on by deci mat i ng
the convol ved i mage, and convol ves t hi s r educed
i mages wi t h a smal ! LOG mas k. I n oder t o r ecover
t he f u l l r es ol ut i on i mage, a i nt er pol at i on i s used
at l as t . The amount of oper at i ons can be r educed by
many t i mes us i ng t hi s f as t al gor i t hm f or a l ar ge
6 , but f or a smal l er 6 , t he decr easi ng of t he
amount of oper at i ons i s not si gni f i cant .
3 Implementation of LOG f i l t er i ng
using Look-Up-Table
I mpl ement at i on of LOG f i l t er i ng usi ng LUT i s an
easy. f as t and ef f ect i ve met hod. The ci r cul ar l y
symmet r i c char act er i s t i c of LOG f unct i on makes t he
si ze of a l ook- up- t abl e accept abl e , Thi s assur es
t he f eas i bi l i t y of our met hod pr esent ed her e.
3. 1
Set- up of discrete LOG masks
The discrete LOG masks ar e set up using,
fallowing fomula:
W h e r e k is chosen to as s ure that the eum of mask
coeficients is near to rero.This means k is chosen
so a s to
1) / 2 (N-1) 12 xtyl P t y l
- (N-1) 1 2 J = - (N-1) 1 2 6 1 2 r 6 a
c c (I-k-)exp (- - 1 =o
At the same time,another mask which aims to
form an adaptive threshod is added to the LOG
mask, then a result mask is formed. Using the
result maak,a decision whether a point is on the
edge or not can be made easily. At last, a
coeficient k is chosen for the result mask to
assure that each coeficient of the mask is integer.
A look-up-table corresponding to the NxN mask
for spatial constant 6 is set up with the size
shown below:
Ntl Nt1
2 2
s = ( - t ( - -1)t. . . t l ) X 2 5 6
INtl) (Nt3)
0
X 256
donated as LUT[ 1) (Nt3) /SI [ 2 5 6 1 shown as Table 1.
Each element in the table is the multiple of
the coeficient at location i in the mask shown
as Fig. 3 and the greylevel from 0 to 255.That is:
LUTlil [jl = w(i) X j .
where w(i) is the coeficient at location i in
the mask
Table 1 look-up-table
3. 2 Set - up o f l ook- up- t abl e.
According to the rules commented above,a N X N
mask can be set up corresponding to the spatial
constant 6 , shown as Fig. 2. Because of the
circularly symmetric characteristic of LOG
function, if we divide the whole mask into five
ar e as donated A-E, we I I find that block A and B
ar e symmetic, block C and block ( At B) ar e
symmetric versus x axis,block D and block ( At B)
are symmetric versus point ( 0, 0 ) , block E and
block (AtB) a r e symmetric versus y axis, too.
I
- N -
Fig. 2 LOG mask
w ( 0)
W(l) w(2)
w(3) w(4) w(5)
. . . w(i-1) w(i) w(it1)
...................
...
..........................
................................ w (T)
(Ntl) (Nt3)
a
where T= - 1
Fig.3 coeficients array in a mask
3. 3 LOG c onvol ut i on
After the look- up- table is set up, LOG
convolution became8 operations on r e a l time to
look u p the look-up-table.
The result of convolution at pixel (x,y) in an
image is the accumulation of N X N values from the
look-up-table according to the greylevel value of
each pixel in N X N ar e a in the image where (I, y)
is the center of this area.There are three mapping
l aye r s in this process.First, the mapping exists
between location ( i, j) in the N O N area where i, j
=-(N-I)/Z, . . . , (N-1)/2 and (itj), then (i2tja)
is mapped to the location in the mask defined as
above range from 0 to T,These two mapping procees
can be performed using look-up-table , too. In the
end, the value from the look-up-table according to
location(i, j) and the greylevel at piont (xti,ytj)
is obtained .There is no multiplying aperations
in this process.
The hardware to perform this process of
computing convolution can be implemented easily
and rapidly, too.
4 The exper i ment r es ul t s
39
The or i gi nal i mage Lenna was used t o t est our
met hod. we chos e t hr ee di f f er ent spat i al const ant
6 and di f f er ent wi dt h N of mas ks cor r espondi ngl y.
The exper i ment r es ul t s ar e shown i n Tabl e 2 .
F # g B t he filtered image r i t h 6 I 4 Fi g 1 t he I d t e r e d b as e w i t h 6 5 0
The s of t war e was i mpl ement ed i n C and r un
under DOS 5. 0 i n P C 386/ 33
The speed- up of convol ut i ng i s obvi ous I f t hi s
pr ocess i s i mpl ement ed by har dwar e, t he speed- up
wi l l be mor e si gni f i cant
Tabl e 2: Exper i ment r esul t s
6 N*N Ti mel ( 8 ) Ti me2 ( **) Speedup
0. 3 313 0. 8 ( sec. 5. 2 ( sec. 1 6. 5
1. 4 11' 11 4 . 4 ( sec. 45. 3 ( sec. j 10. 3
5. 0 31* 31 8( sec. ) 129. 6( sec) 16. 2
Ti mel (*) t i me cons umi ng usi ng our met hod
Ti me2 ( **) t i me cons umi ng usi ng convent i onal met hod
5 Conlusion
We have pr es ent ed a new t echni que t o comput e
convol ut i on whi ch i s based on t he met hod usi ng
l ook- up- t abl e, I t makes edge det ect i on usi ng LOG
f i l t er become a met hod on r eal t i me espesi al l y
wi t h har dwar e i mpl ement at i on.
' Di gi t al St ep Edges Fr om Zero
Second Di r ect i onal Der i vat i ves' . I EEE Tr ans. on
PAMI , 7: 121- 127.
[3] Rober t M. Har al i ck( l 985) . Di gi t al St ep Edges
Fr om Zer o Cr os s i ng of Second Di r ect i onal
Der i vat i ves.
[41 V. Ber z i ns ( 1984) , Accur acy of Lapl aci an Edge
Oper at or s. Comput . Vi s i on, Gr aphi cs & I mage
Pr oces s i ng, 27: 195- 210.
(51 P. J . Bur t ( 19831. Fas t al gor i t hm f or Est i mat i ng
l ocal i mage pr oper t i es. Comput . Vi si on,
Gr aphi cs & I mage Pr oces s i ng, 21: 368- 382
[ 61 W. Wel l s( 1986) . Ef f i ci ent synt hesi s of Gaussi an
f i l t e s by cas caded uni f or m f i l t er s. I EEE
Tr ans on PAMI , PAMMI - 8: 2 3 4 - 239.
[ 71 D. Mar and H. Hi l dr et h( 1980) . Theor y of edge
det ec i on. Pr oc. Roy. Sco. London , B207: 187- 217.
Cr ossi ng of
I EEE Tr ans . on PAMI , No. 1: 58- 68
Ref er ences:
[ l l J .S.Chen et c. ( 1987) , Fas t Convol ut i on wi t h
Lapl aci an of Gaus s i an Mas ks . I EEE Tr ans. on PAMI,
No. 4: 584- 590.
121 W. E. L. Gr i mson & E.C.Hildreth(1985).Comments on
40