0% found this document useful (0 votes)
107 views12 pages1973 - Haralick, Shanmugan, Dinstein - Textural Features For Image Classification PDF
Uploaded by
Ricardus Anggi PramunendarCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
107 views12 pages1973 - Haralick, Shanmugan, Dinstein - Textural Features For Image Classification PDF
Uploaded by
Ricardus Anggi PramunendarCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 12
Rent by emission BEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS
ERIC Ne etree tgp
cm yt al a on Hos
Textural Features for Image Classification
ROBERT M, HARALICK, K. SHANMUGAM, aso ITSHAK DINSTEIN
lating sets we rn ret oa a, wer
‘tle ap el eno ty mT
{in atin cepa, sn tm the pe ee
‘ciao eke tee ict Kins oe St ae
‘Scrooeks of fen of ss 060 jrcioma ih
Facer oat nda tes ol far Rees Fo
‘Mc Ste ENTS maki mary cnt ee nse
‘Step. We eo ino Sisson
‘elo recov pa (cr et ci yt
{va te econ rene ear pap
Ss an a opine te
{5 perce the otra Spr rth al pee
‘Suleiman pce wale ce, Tae as
{ete atte wily compte earl as roby hae 2
‘sel splay er » ne mrt fbn
ITH THE ADVENT of hissed psec urpoe
Wott sonpates ts oman onc rs
iathenatal or sia poses on pir deta
from images of photopic gui. In tat of ese
Pocauc the pra tfomston reed fet
finn of wo arte (4) Te soap te dil
fom sly Rreinthe put ovate dimes
ea oer Uist te
EES CASH Sate Goma eee ar of
FS ae ona x wn es
Sega aie rt
ina Lives Kape tno tog Ge Beate Se
Sawa Wihe Se Oe nthe Kear a
faray. IPL, = (12.0%) and by = (12) ae the
‘Xand Y spatial domins, then, % 1, ithe set of resale
tion cls and the digtal image Fs a fuction which assigns
some gray-one value Ge (12°) to each and every
resolution cells LL, +. Varios two-dimensional
Analyses ae peeformed on J to achiew spec image
roctatng tasks such a coding, restoration, enbancemen,
{nd clastcation Ip recent years a tremendous aroun of
‘computer procesing of photographs has occurred, with
faces having been developed to process saything fom
serial photopzaphs to photomicrographs [I], [2
This paper Weare concerned with he as of developing
ase of features Tor classifying o categorizing pictorial
ata. The clsifcaton of pittoral data can be done on a
‘esoltion cell bass auch ain ening the cop category
‘fa resolution cel on satelite imagery) or on a Block of
contiguous resolution cel (sch atin tenting the crop
‘eatery of an entie agricultural eld encoding over 8
large numberof resolution eel). The moet dificult sep in
categorizing pictorial information from a large Bock of
resoltion eels is that of defining a act of meaning
features to describe the pictorial information from the Bock
‘of resolution eels. Once these features are defined, image
blocks canbe categorized wring any one of muita of
ptern-ecopniontesnigues
In search for meaningful features for describing pice
torial information, itis oaly natural to look tow the
types of featores which human beings use in inerrting
‘torial information, Spectral, text, and conetia!
features are three fundamental patter elements ied ia
human interpretation of color photographs. Spectal fae
tures describe the average tonal aration in various ands
ofthe visible andor infrared portion ofan dstromagaetic
secirum, whereas textural features contain formation tout the spatial ditnbation of fon variations within
bund. Contextual eturs contain information derived rom
blocks of itr data surrounding th ares being analyze.
When snail inage aren from back and white photographs
axe independently processed by a machine, then texture and
"The concept of tone je based on the varying shades of
fay of resolution cells in a photographic sage, while
texture is concerned wit the spatial (ats) eiatebation
‘of gray tones. Teature and tone are not independent con.
‘ple rather, they Beran inextable relationship f0 one
snotier very much ke the reltonship between particle
Sanda wave. Context, texture, and one are always present in
‘he image, although at ines one property can dominate the
other.
“Texture can be evaluated as beng fine, coarse, or smooth;
rippled, mole, iegulat, o¢ ineated. For example, inthe
Inu wopics, ie texte om radar imagery cane i
save of noaresistn: faegrain sedimentary rocks and
‘unconsolidated sediments, while coare texture can be
inieatve of coarse grained sedimentary rocks. A massive
texture with high-contrast components may be indicative of
‘neous rocks A hummocky texture can be indie of
coded igneous oaks
Texture i an inte property of virus all surfaces—
the pain of wood, the weave of «fabri, the pattern of
‘crops in @ fold, ete. Ie contains Imporant formation
‘about the structural ateangement of surfaces and thee
Fete ary for human observers to recognize and deseribe
in empiia terms texture has ben extemelytlractory to
presse deition end to analysis by digital computers
Since the textural properties of images appear 10 cary
‘elu faformation for decrimiation purposes, i in
portant to develop features for este. We present in this
‘Paper compattionally quick procedure for extracting
textual features of images and ince the wetness of
thet Features for diseriminatng between diferent kinds of
image dts
ary image texture stadis have employed autocorea-
tion functions {3}, power spectra [4] estited fist and
second-order Markov meshes [5], and relative frequsncies
fof various gray ives on the unnermakied image [6
These had some degre of succes, but we Know litle more
bout texture ater nding the rule of thew experiments
than before becabc they did nt try to specially define,
sfaracterze, or mode texture. They ony ied rome general
‘mathemati arsformaton wai assigns numbers to the
twansformed image in a nomspecic way.
‘eset alerpe to extrac terra features have been
limited to developing algorithms for extracting specie
image propertis such as courenet and presence of eds.
‘Many foc algoriime have bun developed and ted on
special imagery. The subjective parameters (uch ae the
Selection of thresholds) associated with the techniques do
‘ot enable them to be generalized to imapery oterthaa
that processed by the asthors. Resnly, Rosenfeld sad hi
ivestigators presented a set of procedures for extracting
some textural properties of pictorial data (7}-[9}. In [7]
Rosenfeld and Troy described & procedure for obtining a
mentee ofthe fexture "coarsenes” of images. Their pro
‘dures were hase om the dilrences between the pray-tone
‘values of adjacet image elements and om the autocorrla-
ton f the image pray-one values In [8] Rosenfeld and
“Thurston gave a procedure fo detecting Boundaries separa
ing regions which difer i texture courenes, In [9] Troy
tal deserted graylevel manipsation procedures which
an be sed for preprocessing of pictures before applying
the alporithns given a [7] 36 [8], Procedures for detecting
textural properties such as ints and dbs hate alo beet
suggested by other investigators (e, for example, [10]
[12). Before applying these procedutes to pictures other
than the ones processed by the authors ofthe respective
‘apers, the investigator has to make 2 choice as 10 the
‘ethed appropriate tothe picture inguin a wel a the
telson of parameters Fr the paticslar method.
We are preseting in this paper a general procedure foe
excactng textural properties of blocks ef mage deta. These
features are calculated in the spatial domi, and. the
‘tail natee of textre is taken into account in our
‘rocedve, which [based om the essumpsion that the tex:
{Ure information in an image I omtined inthe overal
for “average” spatial tetionsip which the gay tones in
the image have to One another. We compute 4st of gray
tone spatia-dependence probabiliy-dstebuton matrices
fora given image block and suzsest a st of 14 textural
featores which can be extracted from each ofthese mtr,
tural charecterstcs ws homageney,gray-tone Ina de
pondences (linear structure, conta, gumber and nature
Sf boundaries present, and the complexity ofthe image Ie
isimportant to noe thatthe numberof operation required
to compte any one of these features is proportional tothe
tmber of resolution cel inthe image block. Ie fOr this
etton tht we ell these features quickly compatable
‘Wie also investigate the usefulness of texcura etures foe
‘ategorzng or easailyng image blocks rom thre dierent
ata sets ranging fom high‘esoltion photomicrographs
1 Towesoution satelite imagery. The acuray of la
‘cation on multlass categorization using textual features
‘Of these data as was he eange of 80-20 poreen
(ar initial perspective of texture and tone is based on
the concept that texture and tone bear an inenricable
restionship to one another. Tone and texture are svays
present in an age, although one property can dominate
{he oter at tne. The base itive perceived relation
‘hips between fone and texture are the fllowing When 9
Ssaiarea pach ofan image bas ite varinion~ie, de
“atation of features of svete gray tone—the dominant
property ofthat azea fone. When small-area path has
wide variation of features of sce gray fone, the
Sominaot property ofthat aren it texture, Crcial to this
‘itintion ae the size of the ama-atea patch, the relive
Sizes of the dierete features, and the number of distin
‘ihable dice features, the urber of ditinguishable
onal direte features decreases, the tonal properies wil ‘predominate. Infact, when the smalls patch i only she
‘Sn of ore reoltion cel 50 that thers only one discrete
feature, the only property present is tone. As the numberof
dltingishale features of isccte iy tone increases
within tbe smallare patch, the terre property wil
‘dominate
‘One imporant property of tone-texture isthe spatial
pater of the resolution cls composing cach dete
Tonal feature, When thee is no spatial pttern and the
‘bay-1one vsiaon between features e wide, fine texte
result. As the spacial patters becomes mote definite and
involves more and more resolution cell a course texture
results. An excellent set of photographs of dieent types
of texture may be found in [13
“The preceding description of texture is ofcourse, «gros
simplifeation and idealization of what acta) ocu.
Discrete tonal features ue really quite frzy i hat they do
not neesriy stand out as entities by themehes. There
fore the texture analysis we suggest concerned with move
eral or macroseopiconcrpa than dcrete onl fede.
The procedure we rupees fOr obsiing the textured
features ofan image ib based on the assrapion that the
texture information on an image is contained inthe overall
or “average” spatil relationship which the gay tone in
the image have to one another. More specie we sal
ume that this feature information is adoqeatelyspciod
bbya set of eaytone spatindependence matrices which ae
computed for various angular telationships and cistances
betwen neighboring resolution ell pie on the image, All
of ove textural festares are derived fom these angular
‘eares-neighborgay-tone sata dependence mates,
ray-Tone Spatial Dependence Matrices
Suppose am image to be analyzed is rectangular and bas
Ny, resolution cel im the horizontal dreston and
reiolution cls inthe vertical dieton. Suppowe tht the
{sy fone appearing in each resolution cll is quantized to
IN, levels Let Ly = (1,2) be the orizotal spacial
Addmain, Ly = (1,2.-.N,} be the vertical spatial domai,
tnd C= {UN} be the set of Ny quantized gray
ones. The st, % Lis the et of resoltion els of the
mage ordered by thee fow-column designations. The image
‘Tan be repeseated asa function vbichasigns some gra)
tone in Gta each resolution ell or pit of coordinates it
Ex Lil Lyx by
‘sn een eomponent of our concept! framework of
texture isa measure, or more precisely, fou closely elated
measure from which ll fou textare atures re derived
“These measures are ara termed angulat acaretrephbor
gzytone spain dependence ties, and 10. deserve
these arays we must emphasize our notion of adjacent or
ter i rire ety
Soe eioeain ttc erate
Fgh Rein it 8 tinal pet eben
‘ors melon ar Tand uc 9 eet lb an a
Societe tele ae ae a
Sn pte today oil een as
nearest-neighbor resolution cells themes. We conser a
resolution e=lexcluding thowe on the periphery of an
‘mage, eto have cght nearest-aighborsesoltion cells
asinFee
‘We assume that the texturecontext information in an
image 1 is contained ia the overall or “average” spatial
{clationship which the gny tones in image have 10 one
‘nother. More specially, we shall assume tat this tex:
furecoatent information is adequately spesiod by the
‘matrix of relative frequenses P, wih which two neighbor
Ing resolution vel separated by distance d ocur on the
mage, one with rey tone andthe othr with gray tone J
‘Such matrices of graytone spatialdependence fequencis
sre a fanction of the angular restionsip. between the
reighboring resolution calls a¢ well at x function of the
distance between ther Fig. ilstate the 21 of all
horizontal neighboring resolution el separated by distance
[LT set, along with the mage gay tones, woul he used
Pendeace mate Formaly, for anger quasi to 45
fetes the unnormaliné frequencies ate defied by
Pid) = HUDAOmM ELL, * Ly)
ye LR m = nlm ah
Mt) = tin) = 7)
PUjdAS) = MMe) el, * Le)
yx Lk —m
Mond),
Het) = ina) =}
PG jad) = HUD ECL, * Led
% (by XL) kml =
Fn 0.) = 1m) =)
PU el1957) = HUAN) Elly La)
Hx bg ema d=
ork
Hk) = iH) =),
where f denotes the numberof elements in the st
dine
4
ork =m
men dan
o
NNole that these matrices are syrametrie: Pfs.) =
PU ida. The distance mete p ipl inthe preceding fests} int
fe Pikeet taebagloes mi
oun Wate penal aa,
id Raa Ra ha
Renal: Bove: fecal aaa
pagal. pared. ba
final: plant: fe
oak Be: Fok ot
So a dns rt hing on
Peo
‘equations can be explicitly defined by
Con) = max (Nk ~ mt
Consider Fig. 3, whioh eprescots a4 x 4 image with
four pray tones, ranging from 0 w 3. Fi. 36) shows the
feral form of any graytone spatial dependence matrix
For example, the clement in the (1) postion of the
distance 1 hoszontl Py mate the tot aumber of ines,
two gray tones of value 2 and 1 cexured horzontally
aujacet to eachother. To determine this number, we cout
the mmberof pairs of rexlution callin Ry soch thatthe
Sst resolation cal of the pair as gray tone 2 and the
second resolution cell of the pair has gray tone I. In
5) we calculate al four distance | gray tons epali
“dependence mati
TT needed, the appropiate frequency normalzations for
the matrices are cally competed. When the ritonship is
nearest horizontal neighbor (d = 1, «= O°), there wil be
2W, — 1 neighboring reslation cel pais on each 04
land there are Ny, rows, providing & total of 2NM, — 1)
‘earest horizontal neighbor pars (ee Fig. 3), When the
Telaooship is nenrestighediagonal neighbor (d-~ I,
= 4S) there willbe 20N, ~ 1) 45° neighboring resoltion
xl pairs foreach row except the Bs, foe wbich there are
fone, and there are N, rows, This provides a total of
BN, — NG, ~ 1) nearest iga-aiagoal neighbor pairs,
By tymmety there wil be 2¥4, 1) neared verti
neighbor pis and 2, ~ 1 ~ 1) nearest ef-iagonal
‘eghbor pair. Aer he numberof neighboring rescation
sell pars R used in computing particular gry-tone
Spatishdependence mix is obtained, the. matrix
‘norma by dividing each entry in te matric by
eis appropiate at this point to comment on the com
puitional aspects of obtaining the gray-ione spatiale
Dendence matrices, The numberof operations required to
process an image wsing our procedure i dry propor.
tonal to the sumber of resletion cell m present in the
‘image. In comparison, the number of operations are of the
fonder of» log nif one wishes to use Fourier or Hadamard
teanaorm to extract texto information. Als, to compute
the entries in the gry-tone spatn-dependence matrices,
‘one neds to keep only two ines of image data in core at &
time, Thus no severe storage consents are impose. For
ak ar dma sw ee
Fife feepontge oy Sm ofa ees
FBlecmaraeee OSs
‘most of the images we have process, the computations
have been cared out on a shall digital computer (PDP
15)20 with 12 Kwords of core and two DEC tape drives),
Teswral Features Exnacted from Gray-Tone Spatial
Dependence Maces
‘Our intial assumption in character image texture is
that all the entre information contain inte rey tone
‘pati dependence matrices. Heace all the textural features
We suagen te exacted ffom thee frayfone spate
pendence matics. The equations which define a set of 14
measures of textual fares are given in Appendix 1
Some of these measures relate to speci textural charac
teristics of the image such as homogeneity, contrast, and
the presence of organized structure within the image. Otber
‘measures characorize the compleity and matte of s)-
{one transitions which oasur in the image. Even though
these festues contain information bout the textural
samacterinice ofthe image, iti hard to identify wich
‘gecie textral characteristic is represented by each of
these features
Fer istrative purpoies, we wil define 3 of the 14
textural features inthis section and explain the senifcance
‘ofthese fextures inter ofthe kind of values they take 0”
{or two images of distinctly diferent textural characteitics
“The features we consider ae
aE ECEY ®
a-BeL aCe) ©
R=
oa, Grossland
Angle ASM Contrast Correlation
& .0128 8075
42 [0080 7
90.0077 5987
135° “0064 “410
Avg. :0087 10259
whet ty dn a a ate the menos and standard dene
ston: ofthe marginal drtributions associated with PUR,
nd Risa normalizing constant,
Tig. 4 shows the cigtal printout of two 6 x 64 image
Docks taken from a satelite picture over the Califor
coasine (NASA ERTS Image no, 1002-8139). The image
Showa in Fig. a) ia epresentative sample of grashans
and Fie. 4) iss epreenative sample of water bode in
the ars. The vals ofthe features fy, fo, and fy obtained
from gray-tone spaadependence mattices for distance
(4 1,are shown below the images in Fle 4
The angular second:moment feature (ASM) fi, i 8
measure of homogeneity of the nage. In a homogenous
limage, such as shown in Fig (0), thee are very few dom
Inant gay-tone tnsitons. Hence the P mati for this
image wil have fewer entries of large magnitude, Foran
image lke the one shown in Fig. 2a, the mtr wil have
sarge numberof smal eaves and hence the ASM feature
(which ete wom of squares ofthe entries) nthe P atric
Wil be smaller. A comparison of the ASM values given
Woter Body
Controst Correlation
2.153 7254
31057 ca7e8
3ina 4846,
319 "4850
21863 ‘5327
»
‘elow the images in Fig. 4 sbows the usefulness ofthe ASM
feature as measure of the homogeneity of the image.
"The contrast feature fy is difleence moment ofthe P
matrix and is + measure ofthe contrast or the amount of
focal variations present im an image, Since thee isa large
amount of local variation present the image af Fi. a)
compared tothe image shown in Fg. (0), the contrast
Feature for the grassland image tas consstelly higher
‘aus compare fo the water-body mae.
‘The corlation feature fy is @ messuc of gry-ione
linearsependences i the image. For both the images
shown in Fi. 4, the corclation feature i Somewhat higher
inthe horizontal (0 direction, along the line of san, The
‘tee body image consists mos of constant grajtone
“ale forthe water ps some ade noise Since the noise
Samples are mosly uncorrelated, tbe eoreation features
for the waterbody image have lower values compared to
the grassland image, Alo, the grasland image has 2 cou.
fiderable amount of linea strutture along 45 he cross
the image, and hence the value ofthe correlation feature is aca a TxA ATM Yo A as
higher along this diection compared tothe vals for 90
and 135° directions. ore examples ofthe sgninnce of
Some ofthe textural features ae presented in Append
“The various fetores which we sugeest areal functions
of distance and angle. The angular dependencies presat
Spec problem. Suppose image 4 bas feats o, 6 ad
oranges 0, 45,9", and 135", espctvely, nod image
‘Bis ential 1 4 excep that Bis rotated 90” wth respect
oA. Then B wil have features e, dg, and B for angles 0°,
45,907, and 138 respectively. Since the extare context of
‘inthe are asthe terre context of B, any desson rule
sing the angula features aja! rust proce the same
results for eda. To guarantee this, we suggest tht the
segulay dependent featares not be urd dre. Tneesd
‘we sues tht two functions of band d tel average
fed range (ohich are invariant under rotation), be weds
inputs tothe case.
TL, Arruicsmons of Toerunat FeATURES 108 IKAGE
‘Cusaricanion
In this setion we present the results of ou sta on
the" usefulness of the textural features for extegorizing
images Thee data ets were ued in our stad. These dats
sels were extracted fiom photomicrographs of diferent
rocks, fom acial photographs of man-made and natural
{Sene,and fom highaltude satelite pictures of he earth
[A brie description of the datasets and classifetion a:
Sal be presented. For further ets the interested reader
is refemed to [16}-(18)
Data Set Deeriton
1) Photomicrographs of Sandtnes:Idenifcation ofthe
‘yer of rocks (andeones) present in an underground
Fervor of crue ol is importan in petroleum production
les; ence we undertook n study to ident ferent
types of rocks from the lextural characterise of their
‘photomicrographs. The data set consisted of 243 image
Mocks of ae ot» Gtand he gray tones of the image were
qual-probailty quantized ito 16 fevels. Deals of the
Slporthm usd to perform this quantization are given fo
‘Appendix Il. There were fve sandstone categorie in the
«data ead samples ofthe photomicrographs ae shown in
Fig. 5. The textual featres for the 243 samples were
calculated ffom distance 1 gray-tone spatiakdependence
2) aerial Photographic Data Se: This data set consisted
‘of 170 images belonging to sight cnlegores Theos 170,
images were manually sampled from 9 x 9.n 120000
serial photography negatives provided by the Army En:
‘ronmental Topographic Laboratories, Fort Balti, Va
[17] Toe digtzation ofthe sampled images wnt done on
the Image Diseriination, Enhancement, and Combination
System (IDECS)PDP. The IDECS ian analog digital ear
realtime processing system that as been developed the
Univeaty of Kansas Center for Reseach, Inc A PDP.
°
°
he of nnn) Deir
Upper Mes Ce
‘sole of sttani
“bower Ries
Js capable of performing various function of image pro:
‘sting [25]. The IDECS is capable of prosesing up to
four image inputs Three fying spo: sanners can input
transparencies up toa maximum sizeof x Si, A vdicon
seanner nceps larger tanspatenies ax well & printed
mages. Mlthinage inputs ar electronically repistred and
‘combined into a format more sable for inferpetation
Procesed images are displayed on a color display unit
andjor a black and white monitor. A 24chanae dak is
‘ynchroized with the system and ted for slorage of
Dprocesied digal images. Eicher manually or under com
puter conto, the IDECS/PDP is able to perform image
fnbancement, video level slicing, area integration, extegory
‘dscininaion, display of hsiopams and. scxograma,
‘quantizing, and similar image-processing operations. The
‘lgonthm ised Tor quantizing the serial photographic data
Set digitized the images into eight evel in sucha way that
the areas ofthe difereat levels (n each iste) were os
[Esch image was dived ino a 20 50 aay. The sight
‘tegories were residential old (RSOLD), resents new
(RESNU), LAKE, SWAMP, MARSH, URBAN, RAIL,
and SCRUB or WOOD (SCROD). The fis i categories
‘ontaned 20 samples eac, the RAIL and SCROD eate-
totes contained 10 and #0 samples, respectively. Samples
‘ofthe image are shown in Fig 6
3) Data Set Derived from Satelue Inagery: A vatiety
‘of stl imagery is eutrenty used for the remote censing
of Earth resouees. The most commoaly ued satelite °
‘Rest SE ORR A OSCAR
Hat opeharst SekOb) OOP” ECAUB and wood
‘ager isthe multispectral scanner (MSS) imagery, which
exsist of set of images of «ssn, where cach image in
the set is created By seaming the radiance of the sxne
from the stall in narrow hand of wavelengths The
IMSS operates i waveleagas ranging frm Q3 to 16 xm,
nd upto 24 scanners have been use in the pas.
“The data set used nour study was derived from high:
stude fourchannet multispectral satelite image takes
‘over the California coastine. The image fv one ofthe four
MSS bands is shown in Fig 7. The ground area covered by
the image is about 14400 mi? andthe se ofthe digital
image i 2340 x 5200, Out ofthis are image, a suet of
1624 contiguous image blocks of size 64 x 64 was taken
and used in ou stay. There were even land-se categories
in the image; thet are coastal forest, woodlands, anal
saslands, urban areas, sll and largeiniatd fells and
tater bodies. The land. category for eich Gf © 6
image lock was obtained by interpreting the color com
Prsite ofthe moldspctal image set. The texters etares
were computed from the distance 1 prayfone spi
‘pendence matric of the second MSS band image fn the
rulspetal image set
Sine the ground truth category information was derived
fiom the color composite image, we used st of spectral
(color) features im addition to the textural features for
tategorizing these ages. The spectral features conited
ofthe mean and standard deviation of the gry-tone valves
Of the fx 64 image blocks in etch of the four special
bands
Classiftion Aloricons
1 Piecewise Linear Discriminant Paton Method: For
‘categorizing image blocks in data set Land 3, which have
1 reasonably large numberof simples, we wed pesewise
linearalscriminan fonction method, The data sis were
pictitioned arbitrary Into tuning and text set, The
asian algrth was developed using the samples
{he traning set andi wasted on the samples inthe test
‘A vate of algorithms have been discussed inthe pattern
recognition erature for abtaining pesewise linear dis
‘iminast Function for pattern casieaton [19}-23) tm
4 widely used lgoride, the pattern space i paliioned
Jntoa number of region sing set of hyperplanes Getison
‘oundstes) whose locations ae determined by the sample
patterns. Each ron is dominated by sample ptteras of
Dticula etepory. When a sew pattern fe preset for
‘Sentication, iti assigned a category depending on the
region fo which it belongs Ifthe new pattern Xf located in
‘region dominate by simple patterns of eategory then
{is hse as coming from category
For the muliatgory problem involving Np categorie,
total of VaiMy ~ 12 bypeplanes ate used to pasion
the pattern space. These yperpianes are defined by se
Of ‘weight WeCIOrS Wi f= 12M f= bre,
> fe whic separates the sample pater belonging tothe
‘Rh and th catepories A regresion type algorithm given ia
19. ch] was wsed to obitin the weight vectors, Ar the
lociton ofthe hyperpines are determined, the elses
tion of new patterns i de a follows. For each category
f the number of hyperplanes ¥; which give a postive
response when the new pattern is presented ate dete:
‘ined eng
®
ae where Z's the aupmented pater vector obtined by adding
component of alee 110 Khe,
1
aL]
is send to category cit Y= max (Ft ina
Betoee categories cq ad then assigned 1 cy
Mah? > O,ort0 Gf Wor2 <0, Sever mosbeons
fhe linea dria inton method and multe
of otersaseation procedures may Be ound nth
‘sirens ced
2) Mir ex Deon Rule: For the ssa photograph
data st wich had «seal aumber of amples peters
‘uss min-max decison ole for sategoring the mages
sed on thei tetra fetes. The destin hours
ee bli by ung albu on of the samp inthe
Sata et andthe sample With ase out wan he l=
ti The procedure as repeated forall he spe
the data ett bai he oeral acuta f the casa.
ton sche Tha mod of ang one sample ot”
iely ed by inventor in he paternecognton et
when th data ct has a small sumer of sample,‘
Asin rule dete a flows
‘Delon re I hy and ie the minimum and
o
maxima ales ofthe iiform distribution (=)
fs assigned to category if and ony if
Be SHS tye 91D
m
rs) °
forall j wach that By, 5 dyy = Leo AT thee
exists 0K such tha By 2 sy 2 ayy @ = LN hen
(ese so) i assigned to category kifand aly iF
$ pin te — eal ~ b
F min f
& a= hy
2 Kino tle= poy
& eo 1k @
were Kis the numberof categorie.
Resales of Image Classification States
1) Photomicrograph Data Se Asef 8 variables, com=
‘rising the mean and variance ofthe textural tures fo,
ay a8 foe denitons of thee features, see Append Th
‘ompoted from distance 1 yray-one sptia-dependence
acs, was used as an input to the clase. The data
set was arbitral civiged into a trxning et of 143 samples
and att et of 100 samples. Pisewie linear discriminant
anetions for separating the patterns bloaging tothe five
Sansone categories were dtived using the samples inthe
taining st, The contingscy table for classtying the
samples in the test set shown ia Table L. The overall
axcuacy of the dssication scheme on the test se (Ob-
{ained by summing the diagonal elements in the con-
Yingeney table and ving bythe total mame of samples
sn the tes se) was 89 percent.
rane
conrmciney Fats rerio rorneRSCRAN
ag osama in = 00 nae fa ane
Commonney Tate ret he Ganesan oF 1 A
iota Dats Se “
a
al! 70m opr of hemes er coresy
2) Aerial Photographic Data Ser: A mince decison
rue wis used forthe clnsieaton of 170 mage into ight
Cazgorcs The procesing was dose as follows. Fouegeay-
tone spatal-dependeacics (or four diestions) were som-
puted foreach image. Even textural Features (frfin
etn in Appendix 1) wee computed yeKing a measue-
ment vecor of 4% IL = 44 components for cach image
‘Computation of the men, range and deviation of ach
{ature over the four drstions reduced the ditensonality
10 33. The miimum and maximum statsics were computed
or 169 imapes and the in-man desion role was eed 0
slassly the image Which wan lef! out This procedure was
‘repeated 170 times to ces all he images. Out of thee
170 images, 140, oF 82:3 percent ofthe images, were correaly
laid. The contingency table given in Table I
53) Sullte Imagery Data Set: The pieewise linear-
