Signal Pr ocessi ng 11 ( 1986) 61- 79 61

Nor t h- Hol l and

LOCAL LI NEAR TRANSFORMS FOR TEXTURE MEASUREMENTS
Mi chael UNSER* ( Member EURASI P)
Laboratoire Traitement des Signaux, Ecole Polytechnique F~d~rale de Lausanne, 16 Chemin de Bellerive, CH-1007 Lausanne,
Switzerland
Recei ved 18 Sept ember 1984
Revi sed 3 Sept ember 1985
Abstract. The Nt h or der pr obabi l i t y densi t y f unct i on for pi xel s in a rest ri ct ed nei ghbor hood may be char act er i zed by a set
of N hi st ogr ams ( or s ome cor r es pondi ng moment s ) comput ed al ong appr opr i at el y chos en axes. The pr oj ect i ons on t hose
axes are obt ai ned f r om a l ocal l i near t r ansf or m of t he l ocal nei ghbor hood vect or. This appr oach is cl osel y rel at ed t o filter
bank anal ysi s met hods and gives a statistical j ust i fi cat i on for t he ext r act i on of t ext ure pr oper t i es by means of convol ut i on
oper at or s or l ocal mat ches. Opt i mal and sub- opt i mal l i near oper at or s are der i ved for t ext ure anal ysi s and cl assi fi cat i on.
Exper i ment al resul t s i ndi cat e t hat t he met hod is r obust , flexible, and t hat it per f or ms as well as s t andar d co- occur r ence bas ed
me t hods f or t ext ur e cl assi fi cat i on. The pr opos e d appr oach enabl es t ext ur e char act er i zat i on wi t h a l ower number of f eat ur es
and it is al so comput at i onal l y mor e appeal i ng.
Zusammeafassung. Das Ver f ahr en der Bi l danal yse mi t Hi l fe l okal er l i nearer Tr ans f or mat i onen er l aubt es, di e N- di me ns i ona l e
Ver t ei l ungsdi cht ef unkt i on der Punkt e ei nes begr enzt en Bi l dausschni t t s dur ch N Hi st ogr amme anzun~ihern, di e ent l ang geei gnet
gew~ihlter Achsen aufgest el l t wer den. Di e Pr oj ekt i onen auf di ese Achs en wer den mit Hi l fe ei ner l i near en Tr ans f or mat i on des
s ogenannt en Nachbar s chaf t s vekt or s ber echnet . Di ese Ann~iherung ent spr i cht der Anal yse mit Hi l fe ei ner Fi l t er bank und gi bt
ei ne st at i st i sche Recht f er t i gung fiir di e Ext rakt i on yon Ei genschaf t en der Bi l dt ext ur mi t Hi l fe l okal er Merkmal sfi l t er. Opt i mal e
und s ubopt i mal e Lr s unge n fiir di e Wahl der ar t i ger l i nearer Filter wer den vor geschl agen fiir Text ur anal yse and Kl assi fi kat i on.
Wie Exper i ment e mi t r eal en Bi l dt ext uren zei gen, ist di e Me t hode unempf i ndl i ch, anpassbar , und ebens o zuverl/issig wie
St a nda r dme t hode n di e Pi xel paar ebezi ehungen fiir Text urkl assi fi kat i on ver wenden. Die vor geschl agene Me t hode er mr gl i cht
Text ur bes chr ei bungen mi t weni ger Par amet er n und benr t i gt weni ger Comput er oper at i onen.
R~u mr . La mr t hode d' anal ys e de t ext ur e par t r ans f or mat i on l i nf ai r e l ocal e per met une car act f r i sat i on part i el l e d' une densi t 6
de pr obabi l i t 6 d' or dr e N par N hi st ogr ammes cal cul r s sel on des axes convenabl ement choisis. Cet t e a ppr oc he est 6qui val ent e
une anal yse par banc de filtres et appor t e une j ust i fi cat i on st at i st i que quant au pr i nci pe de l ' ext r act i on de pr opr i r t r s de
t ext ure par des masques de convol ut i on. Des sol ut i ons opt i mal es et sous- opt i mal es pour le choi x des filtres sont pr opos r e s
pour l ' anal yse et la cl assi fi cat i on de t ext ures. On mont r e expr r i ment al ement que la mr t hode est r obust e et flexible. Pour la
cl assi fi cat i on de t ext ures, elle per met d' at t ei ndr e des per f or mances aussi bonne s que les m&hodes usuel l es se bas ant sur des
mesur es de co- occur r ences. De pl us, elle donne l i eu/ l une car act r r i sat i on des pr opr i r t r s de t ext ure avec un nombr e moi ndr e
d' at t r i but s; elle se pr at e 6gal ement h u n calcul pl us ai sr.
Keywords. Text ure analysis~ t ext ur e cl assi fi cat i on, l i near t r ansf or m, l i near oper at or s, filter bank, Ka r h u n e n - Lo r v e t r ansf or m.
1. Introduction
Text ur e is t he t erm used to qual i fy t he surface
of a given obj ect or phe nome non and is undoub-
t edl y one of t he mai n feat ures used in i mage pro-
* Pr esent affiliation: Bi omedi cal Engi neer i ng and
I ns t r ument at i on Br anch, Room 3W13, Bui l di ng 13, Nat i onal
Inst i t ut es o f Heal t h, Bet hesda, MD 20892, U. S. A.
cessing and pat t er n recogni t i on. Text ur e must be
r egar ded as a nei ghbor hood pr oper t y of an i mage
poi nt and has been wi del y st udi ed duri ng t he last
decade. A number of analysis met hods whi ch
all t ry in some way t o descri be pi xel nei ghbor-
hood rel at i onshi ps have been suggested [16]. A
fruitful appr oach, whi ch has been st udi ed by
different aut hors, is to ext ract local nei ghbor hood
0165-1684/ 86/ $3. 50 (~) 1986, El sevi er Sci ence Publ i sher s B.V. ( Nor t h- Hol l and)
62
i nf or mat i on by means of l i near filtering oper at or s
[1, 2, 11, 15, 24, 25, 31]. Thus, t he local t ext ure
propert i es of a given regi on in an i mage can be
char act er i zed by a set of energy measures com-
put ed at t he out put of a filter bank; these convo-
l ut i on masks may be local det ect ors of el ement ar y
st ruct ures such as ' fiats' , ' spot s' , or ' edges' [24, 25],
di rect i onal filters whi ch at t empt to model the
human visual system [ 11, 15, 31 ] or l i near oper at or s
obt ai ned f r om a pri nci pal component analysis of
t he under l yi ng t ext ure field (' eigen-filters' ) [1, 2].
This work present s a statistical vect or space for-
mul at i on of this pr obl em whi ch leads to an
ext ensi on and a uni fi cat i on of previ ous r epor t ed
appr oaches. It suggests represent i ng t he Nt h or der
probabi l i t y densi t y f unct i on ( PDF) of t he pixels
in a rest ri ct ed nei ghbor hood by a set of N his-
t ograms eval uat ed aft er sui t abl e l i near t r ansf or ma-
tion. The t r ansf or m coefficients are comput ed over
a slidhag wi ndow and t he resulting st ruct ure is
equi val ent to filtering t he i mage with a bank of
filters. The choi ce of a statistical ar gument at i on
enabl es t he defi ni t i on of opt i mal and sub-opt i mal
l i near oper at or s f or t ext ur e analysis and
classification. This new f or mul at i on allows t he
compar i son of vari ous t r ansf or ms (i ncl udi ng
Laws' set of oper at or s [24]) in t hei r ability to
ext ract val uabl e t ext ur e i nf or mat i on.
An i mpor t ant aspect of this research is t he
exper i ment al eval uat i on of t he pr oposed met hod,
whi ch is pr esent ed in Sect i on 4. First of all, it
appears t hat t he per f or mance in t ext ure analysis
and classification is not very di fferent among
vari ous sub-opt i mal t ransforms. Second, t he com-
pari son with ot her wel l -known t echni ques i ndi cat e
t hat equi val ent - - and somet i mes i mpr ove d- -
per f or mance in t ext ur e classification can be
obt ai ned usi ng t he statistical i nf or mat i on pr ovi ded
by channel hi st ograms, i nst ead of measures based
on pairs of pixels (e.g., cor r el at i on coefficients,
co- occur r ence mat ri ces). Concer ni ng this last
aspect , however, we quest i on t he exper i ment al
pr ocedur es and concl usi ons of some aut hor s who
state t hat "t ext ur e energy measur es" used on t hei r
own are mor e power f ul t han measures based on
Signal Processing
M. Unser / Local t ransf orms f or t ext ure measurement s
pairs of pixels, in t hat all possible pairwise
confi gurat i ons in t he consi der ed nei ghbor hood
were not account ed f or [24, 25].
Anot her i mpor t ant poi nt of t he present formul a-
t i on is t hat it est abl i shed t he link bet ween a filter
bank analysis met hod and t he t r ansf or m t ech-
ni ques f or whi ch a large vari et y of fast al gori t hms
are available. It shoul d be emphasi zed, however,
t hat t he appr oach pr esent ed in this paper differs,
in a number of ways, f r om t he ot her digital trans-
f or m met hods t hat have previ ousl y been appl i ed
to t ext ure analysis [5, 23]. The first and essential
di fference is t hat t he statistics associ at ed with
i ndi vi dual t r ansf or m coefficients are used as tex-
t ure descri pt ors i nst ead of t he coefficients t hem-
selves (or some funct i ons of t he coefficients). The
second di fference concerns t he choi ce of a rela-
tively small domai n to t r ansf or m (t ypi cal l y 3 x 3
or 5 x 5). Finally, t he t r ansf or m coefficients are
comput ed within a sliding wi ndow while in t he
ot her met hods t hey are t ypi cal l y comput ed on non-
over l appi ng square sub-images.
2. Texture description
Most statistical appr oaches to t ext ure analysis
pr ovi de a part i al char act er i zat i on of t he j oi nt prob-
ability of t he pixels in a rest ri ct ed nei ghbor hood.
In this sect i on, t he pr obl em of descri bi ng local
t ext ure pr oper t i es is rest at ed as t hat of finding an
adequat e statistical descri pt i on of t he di st ri but i on
of a r andom vect or vari abl e: t he local neighborhood
vector. This f or mat i on l ogi cal l y i nt roduces a new
met hod f or t ext ure char act er i zat i on based on a
local l i near t r ansf or mat i on of t he l ocal nei ghbor-
hood vect or. The mot i vat i on of such an oper at i on
is t hat it allows access to rel at i vel y compact statis-
tical measures t hat are st rongl y rel at ed to t he struc-
t ure of t he t ext ur e as it will be shown here.
2.1. Local neighborhood vector
A di scret e t ext ure i mage defi ned on a K x L
rect angul ar grid is denot ed by {Xk.t}, ( k =
1, 2 , . . . , K, ! = 1, 2 , . . . , L) and is consi der ed to be
M. Unser / Local transforms for texture measurements
t he real i zat i on of a bi di mens i onal st at i onar y and
ergodi c st ochast i c process. Since t ext ure is a nei gh-
bor hood pr oper t y, it is pr act i cal to consi der the
pi xel s in a gi ven nei ghbor hood i ndex by (k, l) as
t he component s of a l ocal f eat ur e vect or. For si m-
plicity, t he coupl i ng nei ghbor hood is defi ned as a
r ect angul ar Nx x Ny domai n cent ered on (k, l). The
N component s of t he l ocal nei ghbor hood vect or
Xk, l =[ X ] , , . . , Xn] T are t he sequent i al l y or der ed
pi xel s bel ongi ng to t he Nx Ny r ect angul ar
domai n i ndexed by (k, l). Thi s f or mal i s m t rans-
f or ms t he ori gi nal grey level i mage {Xkj} in a mul t i -
vari at e sequence {Xk.t} ( k = 1 , . . . , K, l = 1 . . . . . L)
as i l l ust rat ed in Fig. 1.
~k, 1 = [ X l X 2 . . - XN ] T
! . . . . . . . . . . . i o o
o o o o o o o o o o o o o o o
o o o o o o o o o o o o o o
o o o o o o o o o o o o o o o o
o o o o o o o o o o o o o o
o o o o o o ~ o o o o o
. . . . . . ~ o oj . . . . . . . .
o o o o o o o o o o o o o o o o o
o o o o o o o ~ o o o o o o o o o
o o o o o o o o o o o o o o o o o
o o o o o o o o o o o o e o o o o
k
Fig. 1. Definition of the local neighborhood vector.
Wi t hi n this f r amewor k, mos t commonl y used
t ext ure meas ur ement s may be vi ewed as est i mat es
of statistics associ at ed wi t h t he local nei ghbor hood
vect or. These statistics are usual l y est i mat ed, by
spat i al aver agi ng, f r om a l arge numbe r of realiz-
at i ons of t he l ocal nei ghbor hood obt ai ned f r om
a regi on of homogeneous t ext ur e in an i mage.
These quant i t i es c a n - - a t l east c onc e pt ua l l y- - be
obt ai ned by i nt egr at i on ( or s ummat i on) of the
N- di me ns i ona l pr obabi l i t y densi t y f unct i on p( x) .
For exampl e, t he grey level hi st ogr am pr ovi des an
est i mat e of t he first or der PDFs of t he i ndi vi dual
component s of t he nei ghbor hood vect or. Not e t hat
t hese di st ri but i ons are all t he same as a di rect
consequence of st at i onari t y. Est i mat es of the
63
second or der PDFs associ at ed wi t h di fferent pai rs
of component s of t he nei ghbor hood are t he well-
known co- occur r ence mat r i ces [17]. Because of
st at i onari t y, t he t ot al numbe r of di st i nct second
or der PDFs for a Nx Ny nei ghbor hood r educes
to 2 Nx N, - Nx - Nv. The fact t hat this t ype of
r epr esent at i on measur es all vi sual l y useful t ext ural
i nf or mat i on has been demons t r at ed by Gagal owi cz
[ 14]. Meas ur ement s bas ed on co- occur r ence statis-
tics bet ween pai rs of pi xel s have been used success-
fully in a l arge numbe r of appl i cat i ons. The mai n
comput at i onal dr awback of such met hods is t he
r equi r ement for large me mor y st orage whi ch
depends on t he second power of t he t ot al numbe r
of grey levels.
For pract i cal reasons, it is al most i mpossi bl e to
est i mat e hi gher or der PDFs unl ess a par amet r i c
model is used. The set of second moment s can be
used to const r uct t he spat i al covar i ance mat r i x
whi ch is defi ned as
Cx = E{ ( x - E{x})" ( x - E{x})X}, (1)
and whi ch pr ovi des a sufficient statistic f or a mul t i -
vari at e Gaus s i an model . Thi s quant i t y is cl osel y
rel at ed to t he spat i al aut ocor r el at i on (or covari -
ance) f unct i on of {xkj}. Because of st at i onari t y,
t he component s of t he l ocal nei ghbor hood mean
vect or are all t he same and t he covar i ance mat ri x
exhi bi t s a cl ose- t o- Toepl i t z st ruct ure with onl y
2 N~Ny - Nx - Ny + 1 di fferent entries. Thi s
appr oach to t ext ure char act er i zat i on is mor e or
less equi val ent to aut ocor r el at i on met hods or
Fouri er power spect r um appr oaches [7, 10, 14, 22,
23, 27].
2.2 Local linear transform
An i nt erest i ng al t ernat i ve to t he pr evi ousl y men-
t i oned appr oaches is obt ai ned f r om a l i near t rans-
f or mat i on of t he local nei ghbor hood vect or [28].
A local l i near pr oper t y ext r act or (or l ocal l i near
t r ansf or m) is defi ned by
Yk, t = TN' xk. I, (2)
where TN = I t , , . . . , tN] x is a nonsi ngul ar N x N
square mat ri x. Thi s equat i on may l ead to t he t wo
Vo|. 11, No. 1,July 1986
64
fol l owi ng i nt erpret at i ons. First of all, it can be
i nt er pr et ed as a change of basis in t he original
pixel space: t he yi' s are t he coefficients of t he local
expansi on of t he nei ghbor hood vect or obt ai ned
with N l i nearl y i ndependent signals defi ned f r om
t he col umns of t he i nvert ed mat ri x T~' . When T N
is an or t hogonal mat ri x ( or t hogonal and normal -
ized col umns), this oper at i on cor r esponds to a
r ot at i on or a reflection in t he original pixel space.
Secondl y, t he local l i near t r ansf or mat i on, bei ng
defi ned f or every i ndex (k, l), is equi val ent to an
N channel cor r el at i on ( or convol ut i on) . The
different rows of t he t r ansf or m mat ri x TN define
a set of masks t hat are used to filter t he i nput
t ext ure {Xk.I} pr oduci ng N out put images {y~]}
(i = 1 , . . . , N) . The cor r espondi ng system is depi c-
t ed in Fig. 2; it is equi val ent to a bank of N finite
i mpul se r esponse filters. Ever y channel will ext ract
a par t i cul ar aspect of l ocal t ext ure pr oper t y. As it
will be shown bel ow, t he efficiency of this analysis
met hod will depend on t he choi ce of t he t r ansf or m
mat ri x TN.
M. Unser / Local t ransf orms f or t ext ure measurement s
2.3. A new set of texture measurements
First or der PDFs of t he nei ghbor hood vect or in
t he original basis do not convey any i nf or mat i on
about local st ruct ure or spat i al ar r angement (this
assert i on is suppor t ed by t he fact t hat t he grey
scale of an image can be modi fi ed, by nonl i near
mappi ng, to fit al most any specified hi st ogram
wi t hout i nt ri nsi cal l y changi ng t he local t ext ure
propert i es). In addi t i on, because of st at i onari t y,
t hese quant i t i es are all t he same. Aft er t r ansf or ma-
t i on, t he first or der PDFs of t he component s of
y ~ , bei ng st rongl y affect ed by t he dependence
rel at i onshi ps bet ween pixels, will be not i ceabl y
different f r om one anot her. The statistics f or bot h
r epr esent at i on of t he l ocal nei ghbor hood vect or
are rel at ed to each ot her by t he fol l owi ng rel at i on-
ships,
p r ( y ) = I T, A - ' .p (x -- T p . y ) , ( 3)
Cy = E{ ( y - E{y}) (y - E { y } ) T} = TNCxT T,
(4)
X
k,1
fi l t re I ]
fi l t re n
, , ( 1 )
"k, l
:_ ,,(2)
Jk,l
~ .u(n)
.Jk,l
~ :. ,,(N)
fi l t re N J k , ]
Fi g. 2. B a n k o f N F I R f i l t er s f o r t h e a n a l y s i s o f t e x t u r e b y l o c a l
l i n e a r t r a n s f o r m.
Signal Processing
where px(x) and py(y) are t he j oi nt PDFs, and Cx
and Cy are t he covar i ance mat ri ces of t he nei ghbor-
hood vect or bef or e and aft er t r ansf or mat i on.
Equat i on (4) shows t hat t he vari ances al ong t he
new axes are heavi l y dependent upon t he covari-
ance st ruct ure of t he t ext ur e field. It can t her ef or e
be suspect ed t hat first or der statistics aft er suitable
t r ansf or mat i on must pr ovi de a useful t ext ure
descri pt i on. These quant i t i es, whi ch are est i mat ed
by spatial averagi ng at t he out put of the filter bank,
will result in a simplified char act er i zat i on of an
unknown mul t i di mensi onal PDF by means of its
measur ed pr oj ect i ons on di fferent presel ect ed axes.
The channel histogram cor r espondi ng to t he trans-
f or m coefficient y~ quant i fi ed with No-grey levels
is an estimate of the set of probabi l i t i es {P<q~>}
( q = 1 , . . . , N):
P~q~> = Prob{y, = q}
(i = 1 , . . . , N, q = 1 , . . . , No). (5)
An equi val ent char act er i zat i on in t erms of moment
M. Unser / Local transforms for texture measurements
est i mat es can be used. The t heor et i cal channel
var i ances are defi ned by
o.~ = Var{y,} = E [ ( y , - ElY, I) 2]
=tTcxt~ ( i =l , . . . , N) (6)
and pr ovi de a sufficient descr i pt i on in t he case of
an under l yi ng gaussi an r a ndom field. The corre-
s pondi ng set of meas ur ement s charact eri zes t he
ener gy di st r i but i on at t he out put of the filter bank.
The nor mal i zed and cent er ed pt h moment s are
defi ned by
/ . ~ ' ) = E [ ( y , - E[y, ])' ]/ o. p. (7)
Of pr act i cal i nt erest are t he t hi rd (skewness) and
f our t h ( Kur t osi s) mome nt s whi ch r espect i vel y pro-
vi de useful measur es of t he a mount of skewness
and peakednes s in a di st ri but i on.
3. Transform selection
Local t ext ur e pr oper t i es may be ext r act ed usi ng
wel l - known t r ans f or ms such as t he di scret e sine
( DST) , cosi ne ( DCT) , Ha d a ma r d ( DHT) or
Ka r hune n- Lo~ve ( KLT) t r ansf or ms. It will be seen
here t hat t he sel ect i on of a gi ven t r ans f or m can be
rel at ed to t he nei ghbor hood statistics of t he t ex-
t ures to be t reat ed. Two di fferent pr obl ems f or
whi ch opt i mal sol ut i ons will be deri ved are con-
si dered next . The first one is concer ned wi t h t ext ur e
anal ysi s wher e one wi shes to find a r epr esent at i on
t hat descri bes as well as possi bl e t he l ocal t ext ure
pr oper t i es. Fr om a st at i st i cal vi ew poi nt , this is
equi val ent in choosi ng t he set of first or der statistics
t hat pr ovi de t he ' bes t ' char act er i zat i on of t he Nt h
or der PDF of t he l ocal nei ghbor hood vect or. The
second pr obl e m is concer ned wi t h t ext ure
cl assi fi cat i on where t he obj ect i ve is t o defi ne a set
of oper at or s t hat can di scr i mi nat e as well as pos-
sible bet ween vari ous t ext ur e fields. Of pr act i cal
i nt erest is t he exi st ence of s ub- opt i mal t r ansf or ms,
whi ch pr ovi de sat i sf act or y resul t s f or bot h of t hese
pr obl ems.
65
3.1. Optimal transform f or texture analysis
The sel ect i on of a t r ans f or m t hat is t he mos t
sui t ed f or this pr obl em has to be eval uat ed wi t h
respect to s ome per f or mance cri t eri on. I n t hi s sec-
t i on, t he di scussi on is rest ri ct ed to t he class of
ener gy pr eser vi ng t r ans f or ms sat i sfyi ng t he con-
dition:
tr( Cy = TN" Cx" T T) = t r(Cx) = No. 2, (8)
where o .2 is t he var i ance of t he t ext ure {Xk,~}. TWO
di fferent aspect s of opt i mal i t y in t he cont ext of
t ext ure anal ysi s are consi der ed next.
Entropy criterion: As has been ment i oned
bef or e, t he first or der st at i st i cs associ at ed wi t h t he
component s of t he nei ghbor hood vect or, expr essed
in t he usual basi s, are all t he same and t her ef or e
rel at i vel y uni nt erest i ng. Thus, we shoul d choose
t he t r ans f or m whi ch pr oduces first or der statistics
as ' di f f er ent ' as possi bl e. Thi s sol ut i on, whi ch has
the great est di fference in var i ance di st ri but i on, is
t he one t hat mi ni mi zes t he ent r opy cri t eri on [30]
N
H( TN) = - Y. 7, l og( y, ) , (9)
i = 1
where t he y, ' s measur e t he rel at i ve channel ener gy
cont r i but i ons:
y, = Var{y~}/tr(C~) = o',2./No. 2
wi t h 0 <~ y~ <~ 1.
I t al so fol l ows f r om t he wel l - known pr oper t i es of
ent r opy measur es t hat H( TN) will be ma x i mu m
when t he 7ds are all equal ; t hi s least f avor abl e
case cor r es ponds to t he uni t y t r ans f or m or any of
its per mut at i ons: Hmax = H( I N) = l og{N}.
Energy criterion: Anot her a ppr oa c h woul d be
to choose t he t r ans f or m such t hat it pr oduces non-
cor r el at ed vari abl es. Thi s condi t i on is fulfilled
when t he fol l owi ng ener gy cri t eri on is maxi mi zed
and equal t o one:
N
E( TN) = Y~ var{y,}2/llCxll 2
i =l
N
= E o .1111G I I (10)
i ~ 1
Vol . 11, No. 1, J ul y 1986
6 6
where I I " II 2 represent s t he Hi l ber t - Schmi dt nor m
(or energy) of a mat ri x and is i nvari ant to any
similarity t r ansf or mat i on. As uncor r el at edness is
a necessary condi t i on---but not always suf f i ci ent - -
f or i ndependence, t he Nt h or der PDF of t he local
nei ghbor hood vect or may be appr oxi mat ed by a
pr oduct of first or der PDFs obt ai ned f r om this
opt i mal represent at i on. Thi s appr oxi mat i on, whi ch
is exact in t he case of a mul t i vari at e Gaussi an
di st ri but i on, has t he same moment s of degree 1
and 2 ( mean and covari ances) and also t he same
proj ect i ons al ong t he pri nci pal axes. On t he ot her
ext reme, cri t eri on (10) is mi ni mi zed in t he initial
basis whi ch cor r esponds t o t he most cor r el at ed
represent at i on, as a di rect consequence of station-
arity.
Both criteria are par t i cul ar cases of a mor e gen-
eral cri t eri on f unct i on t hat has been i nt r oduced in
[29]. Under t he const rai nt of an energy preservi ng
t r ansf or m, it has been shown in [29] t hat t he trans-
f or m t hat opt i mi zes this general cri t eri on is given
by UN = [ U~, . . . , UN] x, wher e t he u~'s are sol ut i on
of t he charact eri st i c equat i on:
Cx" ui = Ai" ui. (11)
The opt i mal basis vect ors are t he ei genvect ors of
t he spatial covar i ance mat r i x Cx and t he opt i mal
vari ances are the cor r espondi ng eigenvalues:
(A,, i = I , . . . , N) . Thus, t he opt i mal t r ansf or m is
the wel l -known Kar hunen- Lo~ve t r ansf or m
( KLT) whi ch enabl es a decor r el at ed dat a rep-
resent at i on. On t he ot her hand, t he worst rep-
resent at i on, with respect to t he general cri t eri on,
is obt ai ned when t he vari ance al ong t he axes are
all t he same, whi ch is preci sel y t he case f or t he
initial r epr esent at i on xk,~ associ at ed to t he i dent i t y
t ransform. An i mpor t ant consequence is t hat any
nont ri vi al t r ansf or m TN will always i mpr ove t he
efficiency when compar ed with t he initial rep-
resent at i on.
The set of eigenfilters obt ai ned f r om t he prin-
cipal axes of t he spat i al covari ance mat ri x has
initially been pr opos ed by Ade [1] f or pri nci pal
component analysis in t he cont ext of t ext ure anal y-
sis. The ability of this set of masks to ext ract t he
Signal Processing
M. Unser / Local transforms f or texture measurements
different aspect s and const i t uent s of a given t ext ur e
have been demonst r at ed in [2].
3.2. Opt i mal t ransf orm f o r t ext ure classification
It is r easonabl e to assume t hat t he most efficient
channel s f or t he di scri mi nat i on bet ween t wo tex-
tures to~ and to2 are t hose in whi ch t he cor r espond-
ing energy cont r i but i ons are as di fferent as poss-
ible. Fol l owi ng this i dea, a per f or mance cri t eri on
associ at ed with a N x N' t r ansf or m mat ri x TN, =
[ t l , . . . , tN,] T with N'<~ N is i nt r oduced as
N '
j = l
N '
2 2 2
= Y. ( o ' j ~ r 2 j + o %/ c r , j - 2 ) ~ 0 , (12)
j = l
wher e cr~j and tr~j ( j = 1 , . . . , N' ) are t he channel
vari ances or energy cont r i but i ons cor r espondi ng
to t ext ur e a~ and to2, respectively. In equat i on
(12), t he t er m cor r espondi ng to t he relative ener gy
cont r i but i on in channel no. j is symmet r i cal - - wi t h
respect to i ndexes 1 and 2 - - a n d i nvari ant to any
scaling f act or of t he cor r espondi ng basis vector.
This quant i t y is mi ni mum and equal t o zero when
t he vari ances o,2 and cr22j are t he same; its i mport -
ance grows as t hei r respect i ve values differ. Not e
t hat f or a given N x N t r ansf or m TN, t he per f or m-
ance cri t eri on (12) is maxi mi zed f or any N' < N
when t he row vect ors t, are i ndexed as
2 2 2 2
trn/o' 21 + tr21/trH/--.
2 2 2 2
or i N / O ' 2 N "-~- Or2N/OrlN.
It is shown in Appendi x A t hat t he opt i mal
t r ansf or m of di mensi on N x N' t hat maxi mi zes t he
cri t eri on f unct i on J~( N' ) , f or N' = I , . . . , N, is
gi ven by UN, = [Ua, , UN'] T wher e t he row vect ors
ul ( i = 1 , . . . , N) satisfy
Cxl" ui = ~/i" Cx2" ui ( i = 1 , . . . , N ) , (13)
where Cxl and Cx2 are t he t ext ure spatial covari-
ance mat ri ces and where t he ei gensol ut i ons of this
equat i on (ui, yi) (i = 1 , . . . , N) are or der ed in t he
M. Unser / Local transforms f or texture measurements
fol l owi ng manner :
y ~ + l / y ~ > y 2 + 1 / y 2 ~ > . . . 1 > y N+ l / y N. (14)
The cor r espondi ng maxi mum val ue of t he cri t eri on
f unct i on J l ( N' ) is
N '
ma x{J ~( S' ) }= ~ ( y , + 1 / %- 2 ) . (15)
i - - 1
A par t i cul ar sol ut i on of equat i on (13) is obt ai ned
with a t r ansf or m mat ri x VN of di mensi on N x N
whi ch si mul t aneousl y di agonal i zes t he t wo spatial
covar i ance mat ri ces Cx~ and Cx2. It is easily
verified, by subst i t ut i on, t hat t he row-vect ors of
such a t r ansf or m satisfy equat i on (13) and t hat t he
scal ar val ues y~ (i = 1 , . . . , N) are t hen det er mi ned
by
y~=A~I~/A~ 2~ ( i = l , . . . , N) , (16)
where A~ ~>and A~ 2> are t he channel vari ances aft er
t r ansf or mat i on f or t ext ur e to, and co2. It can be
demonst r at ed t hat such a nonsi ngul ar t r ansf or m
mat ri x VN exists f or any symmet ri cal mat ri ces Cx~
and C~2. However , this mat ri x is general l y non-
or t hogonal and t her ef or e not uni que. An interest-
ing sol ut i on f or t he si mul t aneous di agonal i zat i on
of t wo covar i ance mat ri ces has been pr opos ed by
Fukunaga and Koont z [13]; it has t he r emar kabl e
pr oper t y of pr oduci ng compl ement ar y ei genval ues
such t hat A~1)+A~ 2)--- 1 (i = 1 , . . . , N) .
3.3. Sub-optimal transforms
The opt i mal sets of masks i nt r oduced in t he
pr ecedi ng sections are pri mari l y of t heoret i cal
value. They depend on t he covar i ance st ruct ure of
t he t ext ures to be anal ysed or classified and are
usual l y cumber some to det ermi ne: bot h sol ut i ons
requi re t he est i mat i on of spatial covar i ance
matrices and t he use of s t andar d- - comput at i onal l y
expens i ve- - numer i cal ei genvect ors ext ract i on
met hods. In t he t ext ur e analysis pr obl em, t he
opt i mal sol ut i on given by t he KLT will be general l y
different f r om one t ext ur e field to anot her. This is
also t rue f or t he sol ut i on gi ven by (13) which, in
addi t i on, is difficult to general i ze when mor e t han
67
two t ext ures have to be consi dered. These
appr oaches are t her ef or e difficult to appl y in prac-
tical appl i cat i ons such as t ext ur e classification or
i mage segment at i on. For t hese reasons, sub-
opt i mal t r ansf or ms shoul d be used whi ch can
consi derabl y simplify t he feat ure ext ract i on pro-
cedure. Anot her advant age is t hat fast al gori t hms
are available for most of t hese t ransforms.
Usually, separabl e t ransforms, such as t he dis-
crete cosine ( DCT) [4], sine ( DST) [19, 20], real
even and odd Four i er ( DREFT and DROFT) [29]
t ransforms, pr ovi de a close appr oxi mat i on of t he
KLT for a wide sense st at i onary process. A di rect
consequence is t hat t hese t r ansf or ms will approxi -
mat el y di agonal i ze t he spatial covari ance mat ri x
of a very large class of t ext ures. It can t her ef or e
be suspect ed t hat t hei r per f or mance in t ext ure
analysis or classification is very close to t he one
t hat woul d be obt ai ned wi t h t he opt i mal sol ut i ons
given by (11) and (13), respectively. Fur t her mor e,
separabl e t ransforms can be comput ed by succes-
sive filtering al ong t he rows and col umns. This is
i l l ust rat ed in Fig. 3, whi ch shows a separabl e filter
bank t hat comput es a r unni ng DST (di scret e sine
t r ansf or m) in a 3 3 nei ghbor hood.
The sets of masks associ at ed with different trans-
forms in t he case of 3 3 are shown in Fig. 4. The
set of masks i nt r oduced by Laws [24] is also rep-
resent ed. It is wort hwhi l e t o not e t he similarity of
this par t i cul ar set of oper at or s and t he sub-opt i mal
DCT or DROFT with whi ch it shares t wo basis
vect ors out of three. This fact is quite surpri si ng
as Laws f ol l owed a qui t e di fferent reasoni ng t han
t he one pr esent ed here and desi gned his filter bank
empi ri cal l y, using a combi nat i on of flat (11 2 11),
spot ( 11- 2 1]) and edge ([1 0 - 1 I) det ect ors; t he
cor r espondi ng basis vect ors also define a local
l i near - - but nonor t hogonal - - t r ans f or m as defi ned
in Sect i on 2.2.
4. Experimental results
This sect i on present s an exper i ment al eval uat i on
of t he pr oposed met hod. The first part is concer ned
VoL 11, No. 1, J ul y 1986
68
X k . I
M. Unser / Local transforms for texture measurements
Y k . I
Y k . I
Y k , I
Y k . I
Fig. 3. Computation of a running 3 x 3 DST by successive row
and column filtering.
with texture analysis and provides an illustration
of the principle of extraction of textural properties
by local linear transformation. The second part
presents a detailed evaluation in the context of
texture classification. In a first step, local texture
properties are characterized by a set of N channel
variances only. The performances of different sub-
optimal transforms are then compared for various
nei ghborhood sizes (2 x 2 up to 5 x 5). The effects
of adding features such as third and fourth
moments are also taken into account. Finally, the
proposed method is shown to compare favorably
with other well-known techniques such as correla-
tion and co-occurrence based methods.
Si gnal P r o c e s s i n g
Fig. 4. Representation of the set of masks for various 3 3
sub-optimal separable transforms. The definitions of the
DREFT (discrete real even Fourier transform), DROFT (dis-
crete real odd Fourier transform) and DEST (discrete even sine
transform) are given in [29].
The 12 Brodatz textures that have been used in
this study and which are displayed in Fig. 5 were
taken from [6]. The original phot ographs were
digitalized and converted into 256x256 picture
arrays quantified into 256 grey levels. Spatial trans-
ducer nonuniformities were compensated by a
local normalization procedure over a 64 x 64 win-
dow [28]. A standard histogram flattening pro-
cedure was then performed producing output
images with 32 equiprobable grey level values. The
experimental data was therefore not distinguish-
able on the basis of first order statistics only.
4.1. Text ure anal ysi s
In these series of experiments a subset of the
initial data (labeled (A), (B), and (C) in Fig. 5)
has been considered more extensively. In a first
step, the spatial covariance matrices were esti-
mated over a 2 x 2 nei ghborhood and the associ-
ated KLT computed. The corresponding masks are
shown in Fig. 6 and appear to be very similar to
those obtained for the various sub-optimal trans-
forms that have been mentioned previously, which
in this case are all equivalent to a 2 x 2 discrete
Hadamard transform (DHT). These sets of masks
M. Unser / L ocal transforms f or texture measurements 69
d
( 1 ) ( 2) ( 3 )
( 4 ) ( s ) = ( A) ( 6 )
( 7 ) ( 8 ) = ( B) ( 9 ) = ( C)
( 1 0 ) ( i l ) ( 1 2 )
Fig. 5, Preprocessed Brodat z t ext ures us ed for experi ment s ( 2 5 6 x 256 pi xe l s wi t h 32 equi probabl e grey l evel s ) .
V o l . 11, NO. 1, J u l y 1 9 8 6
70
|
I)tlT 2x2
I i
KH" 2x2 ( b ) KLT 2x2 ( c )
Fi g. 6. S u b - o p t i ma l a n d o p t i ma l s et s o f ma s k s (2 x 2) f or a n a l y s i s
o f t e x t u r e s ( A) , ( B) , a n d ( C) i n Fi g. 5.
are all f or med f r om a ' flat' det ect or (lowpass filter
usual l y associ at ed with t he first component ) and
vari ous edge det ect ors in t he vertical hori zont al
and di agonal di rect i ons. The images were t hen
filtered by a runni ng 2 x 2 DHT; t he result of this
processi ng f or t ext ure (A) as well as t he channel
hi st ograms are shown in Figs. 7 and 8, respectively.
Also shown are t he channel vari ances and t he
cor r espondi ng eigenvalues. Fr om this exper i ment ,
it appears very cl earl y t hat t he 2 x 2 DHT provi des
an excel l ent subst i t ut e f or t he KLT f or a 2 2
nei ghbor hood. This fact has also been verified by
t he testing of ot her t ext ures. The qual i t y of t he
appr oxi mat i on of t he KLT have been measur ed
by comput i ng t he quant i t i es defi ned in equat i ons
(9) and (10). For bot h ent r opy and energy criteria,
a relative per f or mance measur e has been defi ned as
~' *- ~'(TN)
r/(TN) (17)
~ ' * - ~ ' ( I N) '
where ~*, ~'(TN), and ~'(I:v) are t he cri t eri on values
obt ai ned with t he KLT, t he t r ansf or m mat ri x TN
and t he i dent i t y mat ri x, respect i vel y. In t he case
of 2 x 2 nei ghbor hood, t ypi cal values f r om 99.6%
up to 100% have general l y been obt ai ned.
Various sub-opt i mal t r ansf or ms were also com-
par ed with t he opt i mal KLT f or larger nei ghbor-
Signal Processing
M. Unser / Local transforms for texture measurements
hoods (3 3 up to 5 x 5). Tabl e 1 gives t he results
of this comput at i on in t he case of a 3 3 nei ghbor-
hood f or t ext ures (A), (B), and (C) in Fig. 5; more
det ai l ed results may be f ound in [28]. For t hese
t ext ures, t he DST was always f ound to per f or m
bet t er t han t he ot her t r ansf or ms and nearl y as good
KLT 2x2 ~a) as the opt i mal KLT. A coefficient close to 100%
i ndi cat es t hat t he channel vari ances are as
I di fferent l y di st ri but ed as possible as well as the
t r ansf or m coefficients bei ng nearl y decorrel at ed.
The good per f or mance of t he DST in dat a rep-
resent at i on is not surprising, as it is well known
t hat this t r ansf or m pr ovi des a good appr oxi mat i on
of t he KLT of a separabl e first or der Mar kov when
p/ > 0.5 (see, f or exampl e, [29]), t hat is, when t he
process is essentially highpass.
Ta b l e 1
En t r o p y a n d e n e r g y c r i t e r i o n s c o mp u t e d o n t e x t u r e s ( A) , ( B) ,
I I
a n d ( C) f or v a r i o u s 3 3 l o c a l l i n e a r t r a n s f o r ms
Texture (A)
Transform H(T) h(T) E(T) e(T)
Ident i t y 3 x 3 2.197 0.0% 40.298% 0.0%
DCT ( DREFT) 3 x 3 1.598 91.738% 94.541% 90.856%
DEST (DROFT) 3 x 3 1.676 79.811% 89.035% 81.633%
DST 3 3 1.553 98.559% 99.491% 99.148%
KLT 3 x 3 1.544 100% 100% 100%
Texture (B)
Transform H(T) h(T) E(T) e(T)
Ident i t y 3 x 3 2.197 0.0% 49.994% 0.0%
DCT (DREFT) 3 3 1.805 90,840% 94.878% 89.758%
DEST (DROFT) 3 x 3 1 . 8 5 1 80.182% 90.221% 80.444%
DST 3 3 1.779 97.009% 98.481% 96.963%
KLT 3 x 3 1.766 100% 100% 100%
Texture (C)
Transform H( T) h ( T) E( T) e( T)
Ident i t y 3 x 3 2.197 0.0% 56.634% 0.0%
DCT (DREFT) 3 x 3 1.904 94.730% 97.690% 94.687%
DEST (DROFT) 3 x 3 1 . 9 5 1 79.657% 90.347% 77.724%
DST 3 3 1.893 98.4411% 99.2094% 98.1756%
KLT 3 x 3 1.888 100% 100% 100%
This met hod has also been appl i ed successfully
to t hecont r ol of t ext ur ed surfaces and t he det ect i on
of defect s in textiles [3]. In this last study, it was
f ound t hat sub-opt i mal t r ansf or ms such as t he DST
M. Unser / Local transforms f or texture measurements 71
Fig. 7. Sub- opt i ma l 2 x 2 DHT pr oc e s s i ng of t ext ur e ( C) : l ocal t r a n s f o r m coeffi ci ent i mages .
per f or med as we l l - - a nd somet i mes be t t e r - - t ha n
t he KLT f or t he det ect i on of defect s in mat eri al s.
Very si mi l ar results were obt ai ned wi t h all sets of
masks t hat have been t ri ed; this i ndi cat es t hat t he
met hod is robust .
4.2. Te x t ur e cl assi f i cat i on
For cl assi fi cat i on pur poses, t he t wel ve t ext ures
i mages s hown in Fig. 5 were t hen di vi ded in squar e
regi ons wi t h 50% overl ap. The exper i ment al dat a
set for each class consi st ed of 961 t ext ure sampl es
of di mensi on 16 x 16, 225 sampl es of di mensi on
32 x 32, and 49 sampl es of di mensi on 64 x 64.
An opt i mal Bayesi an deci si on rule was used to
classify t he t ext ure sampl es, based on t hei r
meas ur ed f eat ur e val ues. The class condi t i onal
pr obabi l i t y densi t y f unct i ons of t he M-
di mensi onal feat ure vect or z were assumed to
be mul t i var i at e Gaus s i an di st ri but i ons wi t h mean
vect ors and covar i ance mat ri ces: (#i , Ci) ( i =
1 , . . . , 12). Under such an assumpt i on, t he Bayes
Vol. 11, No. I, July 1986
72
~ l = 14996
0 2 = 14996
M. Unser / Local transforms for texture measurements
/~1 = 12813 ~1 = 12034
0 2 = 1 2 8 0 4 0 2 = 1 2 0 3 3
h 2 = 3 4 1 0
2 = 3 3 8 7
~' 2 = 5 0 7 2
0 . 2 = 5 0 3 6
h 3 = 3 7 4 8
92 = 3 7 4 9
~ 3 = 3 0 5 3
2 = 3 0 7 8
~k 3 = 2529
0 2 = 2 5 6 6
~ 2 = 4137
2 : 4,36
~4 = 6 9 8 I I I ~4 = 1440
o~ 598 ~ o,~:~44~
A
~ 4 ~ 1908
0 2 = 1906
A B C
Fig. 8. Sub-optimal 2 x 2 DHT processing of textures (A), (B), and (C) in Fig. 5: channel histograms.
c l a s s i f i e r , wh i c h mi n i mi z e s t h e t o t a l p r o b a b i l i t y o f
e r r o r , is e q u i v a l e n t t o a s s i g n a t e x t u r e s a mp l e wi t h
f e a t u r e v e c t o r z t o t h e c l a s s wi t h mi n i mu m d i s t a n c e
v a l u e [ 12] :
d, ( z ) = ( z - / z i ) T. C / 1 , ( z - J ~ i ) + l o g { I c , I}
( i = 1 , . . . , 12). ( 18)
F o r e a c h p a t t e r n t h a t h a s b e e n t e s t e d , t h e t r a i n i n g
wa s p e r f o r me d o n t h e r e ma i n i n g s a mp l e s ( ' l e a v i n g -
o n e - o u t me t h o d ' ) u s i n g ma x i mu m l i k e l i h o o d
e s t i ma t e s o f t h e d i s t r i b u t i o n p a r a me t e r s .
4. 2. 1. Cl as s i f i cat i on wi t h c hanne l var i ances
Va r i o u s s u b - o p t i ma l l o c a l t r a n s f o r ms h a v e b e e n
c o mp a r e d i n t h e i r a b i l i t y t o d i s c r i mi n a t e b e t we e n
t h e t we l v e t e x t u r e s i n Fi g. 5. T h e d e f i n i t i o n o f t h e
c o r r e s p o n d i n g b a s i s v e c t o r s ma y b e f o u n d i n [ 20,
26, 29]. T h e s t a t i s t i c s o f N- d i me n s i o n a l l o c a l
n e i g h b o r h o o d v e c t o r we r e c h a r a c t e r i z e d b y a s e t
o f N e s t i ma t e d c h a n n e l v a r i a n c e s . T h e
c l a s s i f i c a t i o n r e s ul t s f o r d i f f e r e n t n e i g h b o r h o o d
s i zes a r e s h o wn i n T a b l e s 2, 3, 4 a n d 5. As c a n b e
s u s p e c t e d , c l a s s i f i c a t i o n b e c o me s mo r e a c c u r a t e as
t h e s i ze o f t h e t e x t u r e s a mp l e s i n c r e a s e s . F o r a
g i v e n s i ze o f t h e l o c a l n e i g h b o r h o o d , t h e p e r f o r m-
a n c e s o f t h e t r a n s f o r ms t h a t h a v e b e e n c o n s i d e r e d
Table 2
Classification of the twelve textures in Fig. 5 using channel
variances estimated after a local linear sub-optimal transform
in a 2 x 2 neighborhood. M is the number of features, Pc the
percentage of correct classification, and N the number of
samples that have been correctly classified
Transform M Size Pc N
DHT 2 x 2 4 16x16 81.86% 9440out ot 11532
DHT 2 x 2 4 32 x 32 94.85% 2561 out of 2700
DHT 2 x 2 4 64 x 64 99.49% 585 out of 588
Table 3
Classification of the twelve textures in Fig. 5 using channel
variances corresponding to various sub-optimal transform in a
3 x 3 neighborhood
Transform M Size Pc N
DST 3 x 3 9 16 x 16 88.66%
DCT 3x3 9 16x16 88.58%
LAWS 3x3 9 16x16 88.45%
DROFT 3x3 9 16x16 88.41%
DST 3 x 3 9 32 x 32 98.26%
DCT 3x3 9 32x32 98.18%
LAWS 3x3 9 32x32 98.18%
DROFT 3x3 9 32x32 98.18%
DST 3x3 9 64x64 100%
DCT 3x3 9 64x64 100%
LAWS 3 x 3 9 64x64 100%
DROFT 3x3 9 64x64 100%
10224 out of 11 532
10216 out of 11 532
10200 out of 11 532
10 195 out of 11 532
2653 out of 2700
2651 out of 2700
2651 out of 2700
2651 out of 2700
588 out of 588
588 out of 588
588 out of 588
588 out of 588
Si gnal Processing
M. Unser / Local transforms for texture measurements
Table 4
Classification of the twelve textures in Fig. 5 using channel
variances corresponding to various sub-optimal transform in a
4 x 4 neighborhood
Transform M Size Pc Nc
DREFT 4 x 4 16 16x16 90.78% 10469 out of 11532
DROFT 4 x 4 16 16x16 90.61% 10 450 out of l1532
DST 4 x 4 16 16 16 90.56% 10444 out of 11532
DCT 44 16 16x16 90.41% 10426 out of 11532
DEST 4 x 4 16 16x16 90.30% 10413 out of 11532
DREFT 4 x 4 16 32x32 99.07% 2675 out of 2700
DCT 44 16 32x32 98.96% 2672 out of 2700
DEST 4 x 4 16 32x32 98.93% 2671 out of 2700
DST 4 x 4 16 32x32 98.77% 2667 out of 2700
DROFT 4 x 4 16 32x32 98.74% 2666 out of 2700
DREFT 4 x 4 16 64x64 100% 588 out of 588
DCT 4 x 4 16 64x64 100% 588 out of 588
DEST 44 16 64x64 100% 588 out of 588
DST 4 x 4 16 6464 100% 588 out of 588
DROFT 4 x 4 16 64x64 100% 588 out of 588
73
a r e v e r y s i mi l a r ; t hi s t e n d s t o i n d i c a t e t h a t t h e
me t h o d i s q u i t e r o b u s t . I n t h e c a s e o f 3 x 3 a n d
5 x 5 n e i g h b o r h o o d s t h e b e s t r e s u l t s a r e o b t a i n e d
wi t h t h e s i ne ( DS T ) a n d c o s i n e ( DC T ) t r a n s f o r ms .
T h e g o o d p e r f o r ma n c e o f t h e 4 4 DR E F T ~ wh i c h
i n t hi s p a r t i c u l a r c a s e i s e q u i v a l e n t t o a 4 x 4
Ha d a ma r d t r a n s f o r m- - i s q u i t e s u r p r i s i n g ; t hi s c a n
b e o f p r a c t i c a l i n t e r e s t b e c a u s e o f t h e a v a i l a b i l i t y
o f v e r y f a s t c o mp u t a t i o n a l g o r i t h ms , T h e d e t a i l e d
r e s ul t s ( c o n f u s i o n ma t r i c e s ) o f t h e c l a s s i f i c a t i o n o f
32 x 32 t e x t u r e s a mp l e s f o r v a r i o u s s i ze o f t h e l o c a l
DS T a r e g i v e n i n T a b l e 6. T h e e f f e c t o f i n c r e a s i n g
t h e s i ze o f t h e n e i g h b o r h o o d is t o i mp r o v e t h e
c l a s s i f i c a t i o n r e s ul t s wi t h o u t i n t r o d u c i n g n e w
e r r or s . I t is a l s o wo r t h wh i l e t o n o t e t h a t
c l a s s i f i c a t i o n e r r o r s u s u a l l y o c c u r i n a s y mme t r i c a l
wa y , b e t we e n t e x t u r e p a i r s wh i c h a r e t h e mo s t
di f f i c ul t t o d i s c r i mi n a t e v i s u a l l y .
Table 5
Classification of the twelve textures in Fig. 5 using channel
variances corresponding to various sub-optimal transform in a
5 x 5 neighborhood
Transform M Size Pc Nc
DST 5x5 25 16x 16 92.03%
DCT 55 25 16x 16 91.96%
DREFT 5x5 25 16x 16 91.94%
DROFT 5x5 25 16x16 91.93%
DEST 5x5 25 16x16 91.78%
LAWS 5x5 25 16x16 91.45%
DST 5x5 25 32x32 99.63%
DREFT 5x5 25 3232 99.63%
DCT 5x5 25 32x32 99.59%
DROFT 5x 5 25 32x32 99.59%
DEST 5x5 25 32x32 99.56%
LAWS 5x5 25 32x32 99.48%
DST 5x5 25 64x64 100%
DREFT 5x5 25 64x64 100%
DCT 5x5 25 64x64 100%
DROFT 5x5 25 64x64 100%
DEST 5x5 25 64x64 100%
LAWS 5x5 25 64x64 100%
10613 out of 11 532
10605 out of 11 532
10602 out of 11 532
10601 out of 11 532
10584 out of 11 532
10546 out of 11 532
2690 out of 2700
2690 out of 2700
2689 out of 2700
2689 out of 2700
2688 out of 2700
2686 out of 2700
588 out of 588
588 out of 588
588 out of 588
588 out of 588
588 out of 588
588 out of 588
4.2.2. Classification wi t h second, third, and
f our t h mome nt s
I n t h e p r e s e n t f o r mu l a t i o n , t h e u s e o f s e c o n d
o r d e r mo me n t s o r c h a n n e l v a r i a n c e s i s o n l y a n
i n t e r e s t i n g a l t e r n a t i v e t o o t h e r we l l - k n o wn t e c h -
n i q u e s wh i c h a r e b a s e d - - - i n a mo r e o r l es s e x p l i c i t
wa y - - o n me a s u r e me n t s t h a t a r e d i r e c t l y r e l a t e d t o
t h e s p a t i a l c o v a r i a n c e ma t r i x . Th i s c a t e g o r y
i n c l u d e s a l l me t h o d s t h a t ma k e u s e o f p o we r
o r e n e r g y me a s u r e me n t s i n t h e s p a t i a l [ 24, 25] o r
s p e c t r a l d o ma i n [ 7] , a s we l l as a u t o r e g r e s s i v e
o r p r e d i c t i v e me t h o d s [ 10, 22, 27]. T h e s e t e c h -
n i q u e s a r e e s p e c i a l l y we l l s u i t e d f o r t h e a n a l y s i s
a n d c l a s s i f i c a t i o n o f Ga u s s i a n p r o c e s s e s .
Un l i k e t h e p r e v i o u s me n t i o n e d a p p r o a c h e s , t h e
p r o p o s e d me t h o d e n a b l e s t h e me a s u r e me n t o f
s o me h i g h e r o r d e r s t a t i s t i c s . A f i n e r d e s c r i p t i o n o f
t h e l o c a l t e x t u r e p r o p e r t i e s ma y b e o b t a i n e d b y
c o mp u t i n g a d d i t i o n a l f e a t u r e s s u c h as t h e es t i -
ma t e d n o r ma l i z e d s k e wn e s s a n d k u r t o s i s d e f i n e d
o n t h e c h a n n e l h i s t o g r a ms ( e q u a t i o n ( 7) ) . T h e
c o r r e s p o n d i n g c l a s s i f i c a t i o n r e s u l t s a r e s h o wn i n
T a b l e 7 i n t h e c a s e o f a 2 x 2 n e i g h b o r h o o d . T h e
a d d i t i o n o f t h e t h i r d o r f o u r t h mo me n t s i mp r o v e s
t h e p e r f o r ma n c e s s i g n i f i c a n t l y . As c a n b e
e x p e c t e d , t h e be s t r e s u l t s a r e o b t a i n e d i n u s i n g
Vol. 11, No. 1, July 1986
74 M. Unser / Local transforms f or texture measurements
Table 6
Confusi on matrices for the Classification of the twelve textures in Fig. 5 using channel vari ances est i mat ed after
a runni ng 2 x 2, 3 x 3, and 4 x 4 discrete sine t ransform (DST)
1 2 3 4 5 6 7 8 9 10 11 12
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
7
8
9
10
11
12
225 0 0 0 0 0 0 0 0 0 0 0
0 217 0 0 0 0 0 0 0 8 0 0
0 2 219 0 0 4 0 0 0 0 0 0
0 0 0 225 0 0 0 0 0 0 0 0
0 0 0 0 215 0 0 0 0 0 10 0
0 0 2 0 0 213 0 0 0 7 0 3
0 0 0 0 0 0 225 0 0 0 0 0
0 0 0 0 0 0 0 225 0 0 0 0
1 0 0 0 0 0 0 1 219 0 0 4
0 4 0 0 0 16 0 0 0 184 0 21
0 0 0 0 26 0 0 0 0 0 199 0
0 0 0 0 0 12 0 0 5 13 0 195
out of 2700. DST 2 x 2: Correct classification 2561
Total score = 94.85%.
1 2 3 4 5 6 7 8 9 10 11 12
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 225 0 0
2 0 224 0
3 0 0 223
4 0 0 0
5 0 0 0
6 0 0 3
7 0 0 0
8 0 0 0
9 0 0 0
10 0 1 0
11 0 0 0
12 0 0 0
DST 3 x 3: Correct classification 2653
Total score = 98.26%.
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 1 0 0
0 0 2 0 0 0 0 0 0
225 0 0 0 0 0 0 0 0
0 225 0 0 0 0 0 0 0
0 0 218 0 0 0 2 0 2
0 0 0 225 0 0 0 0 0
0 0 0 0 225 0 0 0 0
0 0 0 0 0 225 0 0 0
0 0 6 0 0 0 200 0 18
0 0 0 0 0 0 0 225 0
0 0 1 0 0 1 10 0 213
out of 2700.
1 2 3 4 5 6 7 8 9 10 11 12
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
10
11
12
1 225
2 0
3 0
4 "0
5 0
6 0
7 0
8 0
0
0
0
0
0 0 0 0 0 0 0 0 0 0 0
224 0 0 0 0 0 0 0 1 0 0
0 225 0 0 0 0 0 0 0 0 0
0 0 225 0 0 0 0 0 0 0 0
0 0 0 225 0 0 0 0 0 0 0
0 1 0 0 223 0 0 0 0 0 1
0 0 0 0 0 225 0 0 0 0 0
0 0 0 0 0 0 225 0 0 0 0
0 0 0 0 0 0 0 224 0 0 1
0 0 0 0 3 0 0 0 205 0 17
0 0 0 0 0 0 0 0 0 225 0
0 0 0 0 0 0 0 1 8 0 216
out of 2700. DST 4 x 4: Correct classification 2667
Total score = 98,77%.
Signal Processing
M. Unser / Local transforms f or texture measurements
Table 7
Classification of the twelve textures in Fig. 5 using channel
variances as well as third and fourth moments estimated after
a local linear sub-optimal transform (DHT) in a 2 x 2 neigh-
borhood
Transform M Size Pc Nc
2nd order moments
DHT 2x2 4 1616 81. 86% 9440 out of 11 532
DHT 2 2 4 32 x 32 94. 85% 2561 out of 2700
DHT 2x2 4 64x64 99.49% 585 out of 588
2nd and 3rd order moments
DHT 2 x 2 8 16 x 16 86. 17% 9937 out of 11 532
DHT 22 8 32x32 98. 33% 2655 out of 2700
DHT 2x2 8 64x64 100% 588 out of 588
2nd and 4th order moments
DHT 2x2 8 16x 16 84. 23% 9713 out of 11 532
DHT 2 x 2 8 32 x 32 98. 04% 2647 out of 2700
DHT 2x2 8 6464 100% 588 out of 588
2nd, 3rd, and 4th order moments
DHT 2x2 12 16x16 87. 63% 10105 out of 11 532
DHT 2x2 12 32x32 99.19% 2678 out of 2700
DHT 2x2 12 64x64 100% 585 out of 588
c o n j o i n t l y s e c o n d , t h i r d , a n d f o u r t h mo me n t s . Thi s
s hows ve r y c l e a r l y t h a t t he t e xt ur e s us e d i n t he s e
e x p e r i me n t s a r e n o n - Ga u s s i a n a n d t ha t , i n s uc h a
cas e, t he r e i s a n a d v a n t a g e i n us i ng h i g h e r o r d e r
s t a t i s t i c a l i n f o r ma t i o n t o i mp r o v e t he r e s ul t s i n
c l a s s i f i c a t i on.
Va r i o u s n o n n o r ma l i z e d mo me n t s ha ve al s o b e e n
t e s t e d on t h e i r own. I t was f o u n d t ha t even
mo me n t s - - t h e set o f c h a n n e l v a r i a n c e s b e i n g t he
mo s t p o we r f u l - - a l wa y s p e r f o r me d b e t t e r t h a n o d d
mo me n t s .
4. 2. 3. Co mp a r i s o n wi t h c o r r e l a t i o n me t h o d s
Eq u a t i o n (6) s hows t h a t t he k n o wl e d g e o f t he
s p a t i a l c o v a r i a n c e ma t r i x e n a b l e s t he d e t e r mi n a -
t i on o f t he p o we r at t he o u t p u t o f a ny F I R f i l t er ,
as l ong as i t s s u p p o r t i s e n t i r e l y c o n t a i n e d i n t he
r e g i o n de f i ni ng t he l oc a l n e i g h b o r h o o d . Thi s i s
al s o t r ue wh e n t he s e q u a n t i t i e s a r e e s t i ma t e d b y
s p a t i a l a ve r a gi ng. Th e r e f o r e , t he s t a t i s t i c a l i nf or -
ma t i o n c o n v e y e d b y t he set o f al l c o r r e l a t i o n
coef f i ci ent s s h o u l d be mo r e c o mp l e t e t h a n t he one
a s s o c i a t e d t o t he c h a n n e l va r i a nc e s . F o r ma l l y ,
t he s e d e s c r i p t i o n s a r e o n l y e q u i v a l e n t wh e n t he
75
l oc a l l i n e a r t r a n s f o r m d i a g o n a l i z e s t he s p a t i a l
c o v a r i a n c e ma t r i x. I t is a l s o i mp o r t a n t t o me n t i o n
t ha t t he c h a n n e l v a r i a n c e s p r o v i d e a mo r e c o mp a c t
d e s c r i p t i o n o f s e c o n d o r d e r s t a t i s t i c a l i n f o r ma t i o n
( N c h a n n e l v a r i a n c e s ve r s us 2 ( N- x / N) c or r e l a -
t i on coef f i ci ent s f or a x / N x ~ n e i g h b o r h o o d ) .
I n ' o r d e r t o me a s u r e t he l oss o f i n f o r ma t i o n d u e
t o s u b - o p t i ma l t r a n s f o r m p r o c e s s i n g , t he t e xt ur e
s a mp l e s ha ve be e n c l a s s i f i e d b a s e d on me a s u r e -
me nt s o f t he c o r r e l a t i o n coef f i ci ent s . The s e qua n-
t i t i es wer e n o r ma l i z e d wi t h r e s p e c t t o t he t e xt ur e
v a r i a n c e ; t hi s l as t f e a t ur e is me a n i n g l e s s f or o u r
p u r p o s e o wi n g t o t he p r e p r o c e s s i n g whi c h i mp o s e d
t he s a me fi rst o r d e r s t a t i s t i c s f or al l t es t i ma ge s .
The s e r es ul t s a r e s h o wn i n Ta b l e 8 a n d mu s t be
c o n s i d e r e d as a u p p e r b o u n d o f wha t c a n be
a c h i e v e d b y me a n s o f s e c o n d o r d e r s t at i s t i cs onl y.
As wo u l d be e x p e c t e d , t he p e r f o r ma n c e s a r e
s l i ght l y b e t t e r t h a n t h o s e b a s e d on me a s u r e me n t s
o f c h a n n e l v a r i a n c e s ( Ta b l e s 2 a n d 3). Ho we v e r ,
t hi s di f f e r e nc e is ver y s ma l l , e s p e c i a l l y f or a 2 2
n e i g h b o r h o o d . I n t he cas e o f 3 3 n e i g h b o r h o o d ,
t he l os s o f p e r f o r ma n c e wh e n us i ng c h a n n e l var i -
a nc e s is o n l y o f t he o r d e r o f 1%, whi c h is a ver y
g o o d r e s ul t i f one t a ke s i n a c c o u n t t ha t ni ne
f e a t ur e s a r e u s e d i n s t e a d o f t wel ve. Thi s d e mo n -
s t r at es e x p e r i me n t a l l y t h a t t he s u b - o p t i ma l t r a ns -
f or ms t ha t ha ve b e e n c o n s i d e r e d ar e n e a r l y as
p o we r f u l as a h y p o t h e t i c a l s o l u t i o n t ha t wo u l d be
o p t i ma l - - i n t he s ens e o f e q u a t i o n ( 1 3 ) - - f o r al l
p o s s i b l e pa i r s o f t e xt ur e s i n Fi g. 5. F u r t h e r mo r e ,
i t a p p e a r s t ha t t he r e s ul t s o b t a i n e d i n Se c t i on 4. 2. 2,
us i ng h i g h e r o r d e r mo me n t s , ar e muc h b e t t e r t h a n
wh a t c o u l d ha ve b e e n o b t a i n e d by me a n s o f s e c o n d
o r d e r s t a t i s t i c a l me a s u r e s onl y.
4. 2. 4. Co mp a r i s o n wi t h c o - o c c u r r e n c e me t h o d s
Sp a t i a l gr e y- l e ve l d e p e n d e n c e or c o - o c c u r r e n c e
ma t r i c e s a r e one o f t he mo s t p o p u l a r a n d p o we r f u l
s our c e s o f f e a t ur e s f or t e x t u r e c h a r a c t e r i z a t i o n
[17]. The s e ma t r i c e s a r e u s u a l l y c o n s i d e r e d as
i n t e r me d i a t e me a s u r e s a n d ar e u s e d t o def i ne mo r e
g l o b a l t e x t u r e f e a t ur e s s uc h as c o r r e l a t i o n , c on-
t r as t , et c. I f one c o n s i d e r s a l oc a l n e i g h b o r h o o d
o f d i me n s i o n N= Nx x Ny, i t i s p o s s i b l e t o
Vol. 11, No. 1, Jul y 1986
76 M. Unser / Local transforms f or texture measurements
Table 8
Classification of the twelve textures in Fig. 5 using the different correlations coefficients associated
to a 2 x 2 and 3 x 3 nei ghborhood
Features M Size Pc Arc
2 x 2 Correl at i on coefficient 4 16 17 81.93%
2 x 2 Correl at i on coefficient 4 32 x 32 96.04%
2 2 Correl at i on coefficient 4 64 x 64 99.49%
3 x 3 Correl at i on coefficient 12 16 x 16 89.65%
3 x 3 Correl at i on coefficient 12 32 x 32 99.37%
3 x 3 Correl at i on coefficient 12 64 x 64 100%
9448 out of 11 532
2593 out of 2700
585 out of 588
10 338 out of 11 532
2683 out of 2700
588 out of 588
comput e 2 Nx Ny - Nx - Ny di fferent co- occur r ence
mat ri ces, each of t hem cor r es pondi ng to a gi ven
rel at i ve di spl acement , whi ch is usual l y speci fi ed
by a di st ance val ue d and an or i ent at i on 0. These
mat ri ces pr ovi de est i mat es of t he j oi nt pr obabi l i t y
densi t y f unct i ons f or all di fferent pai rs of pi xel s
in t he consi der ed domai n. Obvi ousl y, this t ype of
statistical char act er i zat i on is mor e power f ul t han
t he cor r es pondi ng set of cor r el at i on coefficients.
The mai n dr awback of t hi s t ype of a ppr oa c h is its
r equi r ement f or excessi ve me mor y st orage and
comput at i on time.
In this st udy, we have rest ri ct ed oursel ves to a
2 x 2 nei ghbor hood. The f our cor r es pondi ng sym-
met ri c co- occur r ence mat r i ces (d = 1 and 0 = 0, 45,
90, 135 degrees) have been c omput e d on every
t ext ure s ampl e and used to eval uat e t wo di fferent
sets of feat ures. The first set consi st s of t he most
c ommonl y used feat ures, namel y Har al i ck et al . ' s
' ener gy' , ' ent r opy' , ' cor r el at i on' , and ' i ner t i a' [17].
The second set of feat ures, namel y, ' cl ust er shade' ,
' cl ust er pr omi nence' , ' homogenei t y' , ' i ner t i a' , and
' cor r el at i on' , has been pr opos e d mor e recent l y by
Conner s et al. [8, 9]. As a mat t er of fact, bot h of
t hese t ext ur e descr i pt i ons are mor e compl et e t han
t he one t hat woul d be obt ai ned measur i ng second
moment s onl y. The resul t s in t ext ure cl assi fi cat i on
are gi ven in Tabl e 9 and have to be c ompa r e d wi t h
t he val ues in Tabl e 7 obt ai ned by l ocal l i near
t r ansf or mat i on. It is not sur pr i si ng t hat co- occur -
rence feat ures are mor e power f ul t han cor r el at i on
or t ext ure ener gy measur es al one. However , it is
wor t hwhi l e not i ci ng t hat t he per cent ages of correct
cl assi fi cat i on obt ai ned wi t h t he 2 x 2 DHT, usi ng
second, t hi rd, and f our t h moment s , compar e favor-
abl y wi t h t hose achi eved wi t h set no. 1. Thi s is
even mor e r emar kabl e since t he numbe r of
f e a t ur e s - - a s well as t he comput at i onal ef f or t - - i s
Table 9
Classification of the twelve textures
co-occurrence matrices defined in a
in Fig. 5 using co-occurrence features comput ed from the four
2 x 2 nei ghborhood
Features M Size Pc Nc
Set 1 (d = 1, 0 = 0, 45, 90, 135 ) 16 16 16 88.82%
Set 1 (d = 1, 0 = 0, 45, 90, 135 ) 16 32 x 32 97.74%
Set 1 (d = 1, # =0, 45, 90, 135 ) 16 6464 99.83%
Set 2 (d = 1, 0 =0, 45, 90, 135 ) 24 16x 16 91.19%
Set 2 (d = 1, 0 = 0, 45, 90, 135 ) 24 32 x 32 99.63%
Set 2 (d = 1, # = 0, 45, 90, 135") 24 64 64 100%
10243 out of 11 532
2639 out of 2700
587 out of 588
10516 out of 11 532
2690 out of 2700
588 out of 588
Signal Processing
M. Unser / Local transforms for texture measurements
significantly less i mpor t ant in t he first case. The
excel l ent results of set no. 2 are not surpri si ng i f
one takes in account t he effort t hat many research-
ers have spent, duri ng t he last decade, t o i mpr ove
feat ure ext, ' action based on co- occur r ence
matrices. It is bel i eved t hat a similar i mpr ovement
coul d be achi eved wi t h an opt i mi zed set of
measur ement s based on a l i near t r ansf or mat i on of
t he local nei ghbor hood vect or.
Fr om t hese last exper i ment s, it can be concl uded
t hat t he pr oposed appr oach, pr ovi di ng t hat hi gher
moment s are comput ed, is al most as efficient for
t ext ure di scri mi nat i on as met hods based on co-
occur r ence measurement s. Mor eover , it seems pre-
ferabl e, in pract i cal appl i cat i ons, to ext ract local
t ext ure pr oper t i es usi ng l ocal l i near t ransforms
r at her t han co- occur r ence matrices, f or t he follow-
ing reasons:
(i) It enabl es a mor e compact descri pt i on of
local t ext ur e propert i es. It is also easy to adapt f or
vari ous sizes of t he l ocal nei ghbor hood.
(ii) The eval uat i on of t he channel hi st ograms
or some associ at ed moment s is comput at i onal l y
less demandi ng.
(iii) Because of its paral l el st ruct ure, this
al gori t hm is especi al l y sui t ed f or an i mpl ement a-
t i on on a speci al i zed paral l el image processor; it
is t her ef or e easily appl i cabl e in t he cont ext of real
t i me i mage processi ng.
(iv) The pri nci pl e f or t he ext ract i on of t ext ure
feat ures by l i near filtering is in agr eement with
recent physi ol ogi cal and psychol ogi cal findings
on t he visual system [ 18, 21] . Fur t her mor e,
classification errors general l y occur bet ween tex-
tures t hat are t he most difficult to di fferent i at e
visually; this is not necessari l y t he case when co-
occur r ence measur ement s are used.
5. Co nc l us i o n
It has been shown t hat l ocal l i near t r ansf or ms
can efficiently be appl i ed to t ext ure analysis and
classification. The pr opos ed met hod gives access
to hi gher or der statistical i nf or mat i on by means of
simple hi st ogram or moment comput at i on al ong
77
sel ect ed axes in t he space of pixel values in a
specified nei ghbor hood. Opt i mal sets of masks
have been deri ved f or t ext ur e analysis and t wo
class t ext ur e classification. The pract i cal i mport -
ance of sub-opt i mal t r ansf or m processi ng has been
emphasi zed as it consi der abl y simplifies comput a-
t i on and is appl i cabl e to cases where mor e t han
t wo t ext ures have to be consi der ed ( N class t ext ur e
classification pr obl em or segment at i on).
Fr om exper i ment al eval uat i on and compar i son
with ot her wel l -known t echni ques, t ext ure
measur ement s obt ai ned f r om t he channel his-
t ograms appear to have t he fol l owi ng advant ages
whi ch shoul d make t hem useful in most pract i cal
appl i cat i ons:
Good per f or mance: Channel vari ances al one
were f ound to be al most as power f ul as t he corre-
spondi ng set of correl at i on coefficients. I f hi gher
or der statistics are also ext ract ed, using t hi rd and
f our t h moment s, t he resul t i ng set of feat ures yields
as good a classification as st andar d co- occur r ence
measures based on pairs of pixels.
Compact nes s o f t he descri pt i on: Usually, t he
pr oposed appr oach uses less feat ures t han met hods
based on pairs of pixels. The number of channel
hi st ograms is always smal l er t han t he number of
different correl at i on coefficients or co- occur r ence
mat ri ces t han can be defi ned f or the same neigh-
bor hood (for exampl e, si xt een channel vari ances
vs. t went y f our cor r el at i on coefficients for a 4 4
nei ghbor hood) .
Robus t nes s : Per f or mance in t ext ure analysis
or classification is not great l y affected by t he exact
shape of t he masks whi ch i ndi cat es t hat t he met hod
is robust.
Fl exi bi l i t y: The qual i t y of t he t ext ure descrip-
tion can be i mpr oved, accor di ng to the needs, by
simply i ncreasi ng t he size of t he local neigh-
bor hood.
Si mpl e al gori t hmi c st ruct ure: Very fast and
efficient al gori t hms are avai l abl e for most sub-
opt i mal t r ansf or ms t hat have been consi dered.
In addi t i on, because of its vect ori al form, t he
met hod is especi al l y well sui t ed f or paral l el
i mpl ement at i on.
Vol. 11, No. 1, Jul y 1986
78 M. Unser / Local transforms for texture measurements
Acknowl edgment
The aut hor is grateful to Profs. Murat Kunt ,
Frederick De Coul on and Murray Eden for their
support and valuable suggestions.
Appendi x A. Opt i mal t ransf orm f or two cl ass t ext ure
cl assi f i cat i on
A ge ne r al i z e d f orm o f t he pe r f or manc e cri t eri on
def i ned by e qua t i o n ( 12) wi t h a N x N' ( N ' ~ < N )
t ransform matrix T = [ U l , . . . , UN'] T is given by
N'
J = E 2 2 2 2
G{ o2J trl s + o ' J cr2j}, (A.1)
j =l
where G{x} is an increasing funct i on for x I> 2.
Clearly, the t erm correspondi ng to the t ransformed
coefficient j is mi ni mum for cr~s = tr22s. The more
variances will differ from each other, the easier it
will be to distinguish bet ween textures COl and co2.
From this poi nt of view, the most favorable trans-
form is the one t hat maximizes J ( N' ) .
The opt i mal basis vectors are solutions of
3J/Ouj=O ( j =l , 2 , . . . , n ) . (A.2)
Because of the condi t i on t hat aG{ x} / Ox>O for
x I> 2, the eval uat i on of the partial derivatives of
J ( N) yields:
Cx, . us/ ( u T . G , . " s) = G 2. u J ( u T . Cx2. us)
( j = 1, 2 , . . . , N)
which is equi val ent to
Cx~ " us = ys" Cx2" u s
(A.3)
( j = 1 , 2 , . . . , N) .
(A.4)
This is easily verified in mul t i pl yi ng this expression
by u T. Thus, the value of )'s is given by
~'s = ( uT " G , " u s) / ( u f . cx2. us)
( j = 1 , 2 , . . . , N) , (A.5)
which is in agreement with equat i on ( A. 3) .
The opt i mal cri t eri on val ue is f o u n d by s ubs t i t u-
t i on in (A.1) of the quant i t i es given by equat i ons
Signal Processing
( A. 3) and ( A. 5) :
N'
ma x{J ( N' ) }= Y~ G{39+l / yj }
j =l
N'
= E G{"T' ( c; ~' Cx2
j - I
q' - Cx 1 Cxl ) ' / . / j }, (A. 6)
where the solutions (uj, yj) ( i = 1 , 2 , . . . , N) are
ordered accordi ng to equat i on (14).
Ref erences
[1] F. Ade, "Application of principal component analysis to
the inspection of industrial goods", SPIE Conf., Geneva,
1983.
[2] F. Ade, "Characterization of texture by 'eigenfilter'", Sig-
nal Processing, Vol. 5, No. 5, September 1983, pp. 451-457.
[3] F. Ade, N. Lins and M. Unser, "Comparison of various
filter sets for defect detection in textiles", Proc. 7th Inter-
nat. Conf. on Pattern Recognition, Montrdal, Canada, July
30-August 2, 1984, pp. 428-431.
[4] N. Ahmed, T. Natarajan and K.R. Rao, "Discrete cosine
transform", IEEE Trans. Comput., Vol. C-23, January
1974, pp. 90-93.
[5] R. Bajcsy and L. Lieberman, "Texture gradient as a depth
cue", Computer Graphics and Image Process., Vol. 5, 1976,
pp. 52-67.
[6] P. Brodatz, Textures--A Photographic Album for Artists
and Designers, Dover, New York, 1966.
[7] C.H. Chen, "A study of texture classifications using spec-
tral features", Proc. 6th Internat. Conf. on Pattern Recogni-
tion, M/inchen, 1982, pp. 382-385.
[8] R.W. Conners, "Toward a set of statistical features which
measure visually perceivable qualities of textures", Proc.
Pattern Recognition Image Processing Conf., August 1979,
pp. 382-390.
[9] R.W. Conners and C.A. Harlow, "Towards a structural
textural analyser based on statistical methods", Comput.
Graphics and Image Process., Vol. 12, 1980, pp. 224-256.
[10] P. De Souza, "Texture recognition via autoregression",
Pattern Recognition, Vol. 15, No. 6, 1982, pp. 471-475.
[ 11 ] O.D. Faugeras, "Texture analysis and classification using
a human visual model", Proc. lnternat. Joint Conf. on
Pattern Recognition, Kyoto, Japan, November 1978, pp.
549-559.
[12] K. Fukunaga, Introduction to Statistical Pattern Recogni-
tion, Academic Press, New York, 1972.
[13] K. Fukunaga and W.L. Koontz, "Application of the
Karhunen-Lorve expansion to feature selection and
ordering", IEEE Trans. Comput., Vol. C-19, No. 4, April
1970, pp. 311-318.
[14] A. Gagalowicz, "Vers un modrle de textures", Thrse de
Doctorat des Sciences, Universit6 de Paris VI, France,
1983.
M. Unser / Local transforms f or texture measurements
[15] G.H. Granlund, "Description of texture using the general
operator approach", Proc. 5th Internat. Conf. on Pattern
Recognition, 1980, pp. 776-779.
[16] R.M. Haralick, "Statistical and structural approaches to
texture", Proc. I EEE, Vol. 67, 1979, pp. 786-804.
[17] R.M. Haralick, K. Shanmugan and I. Dinstein, "Textural
features for image classification", I EEE Trans. Systems,
Man and Cybern., Vol. SMC-8, No. 6, November 1973,
pp. 610-621.
[18] D.H. Hubel and T.N. Wiesel, "Brain mechanism of
vision", Scientific American, Vol. 241, No. 3, September
1979, pp. 130-146.
[19] A.K. Jain, "A fast Karhunen-Lo~ve transform for a class
of stochastic processes", I EEE Trans. Commun., Vol.
COM-24, September 1976, pp. 1023-1029.
[20] A.K. Jain, "A sinusoidal family of unitary transforms",
I EEE Trans. on Pattern Anal ysi s and Machine Intelligence,
Vol. PAMI-I, No. 4, October 1979, pp. 356-365.
[21] B. Julesz and R. Bergen, "Textons, the fundamental ele-
ments in preattentive vision and perception of textures",
The Be#Syst em Tech. J., Vol. 62, No. 6, July/August 1983,
pp. 1619-1645.
[22] R.L. Kashyap, R. Chellappa and A. Khotanzad, "'Texture
classification using features derived from random field
models", Pattern Recognition Lett., Vol. 1, 1982, pp. 43-50.
[23] L. Kirvida, "'Texture measurements for automatic
classification of imagery", I EEE Trans. EMC, Vol. 18,
No. 2, 1976, pp. 38-42.
79
[24] K.I. Laws, "Textured image segmentation", Ph.D. Thesis,
Rept. 940, Image Processing Institute, Univ. of Southern
California, January 1980.
[25] M. Pietikai'nen, A. Rosenfeld and L.S. Davis, "'Experi-
ments with texture classification using averages of local
pattern matches", I EEE Trans. on Systems, Man and
Cvbernet., Vol. SMC-13, No. 3, 1983, pp. 421-426.
[26] W.K. Pratt, Digital Image Processing, Wiley--Interscience,
New York, 1978.
[27] C.W. Therrien, "Linear filtering models for texture
classification and segmentation", Proc. 5th lnternat. Joint
Conf on Pattern Recognition and Image Processing, Miami,
1980, pp. 1132-1135.
[28] M. Unser, "Description statistique de textures: Applica-
tion /t rinspection automatique", Ph.D. Thesis No. 534,
Ecole Polytechnique F6d6rale de Lausanne. Switzerland,
1984.
[29] M. Unser, "On the approximation of the Karhunen-Lo~ve
transform of stationary processes", Signal Processing, Vol.
7, No. 3, December 1984, pp. 231-249.
[30] S. Watanabe, "Karhunen-Lo~ve expansion and factor
analysis", Trans. 4th Prague Conf. on lnJbrmation Theory,
1965, Czechoslovak Academy of Sciences, Prague, 1967.
[31] D. Wermser and C.-E. Liedtke, "Texture analysis using a
model of the visual system", 6th Internat. Con[~ on Pattern
Recognition, Mfinich, Germany, October 19-22, 1982,
pp. 1078-1081.
V o l . I 1 , N o , 1 , J u l y 1 9 8 6