© 2025 IEEE. Pe sonal use o his ma e ial is pe mi ed. Pe mission om IEEE mus be ob ained o all o he uses, in any cu en o u u e
media, including ep in ing/ epublishing his ma e ial o ad e ising o p omo ional pu poses, c ea ing new collec i e wo ks, o esale o
edis ibu ion o se e s o lis s, o euse o any copy igh ed componen o his wo k in o he wo ks.
DOI: 10.1109/EEITE65381.2025.11166185
CIEDE2000-based Wine Colo Analysis Using
Sma phones in Uncons ained En i onmen s
Ch is ina Mas alexi1,2 , B uno Scholles Soa es Dias1,2 , Cla a Coelho1,3, Ped o Ca alho1,3
1Cen e o Telecommunica ions and Mul imedia, INESC TEC, Po o, Po ugal
2Facul y o Enginee ing, Uni e si y o Po o (FEUP), Po o, Po ugal
3ISEP, Poly echnic o Po o, Po o, Po ugal
Emails: [email p o ec ed], [email p o ec ed],
[email p o ec ed], ped o.m.ca [email p o ec ed]
Abs ac —Wine colo o e s key insigh s in o composi ion,
o igin, and aging, se ing as a c ucial indica o o classi ica ion
and quali y assessmen . T adi ional me hods, based on senso y
e alua ion and spec opho ome y, a e o en en i onmen ally
sensi i e and expensi e. This s udy explo es digi al image analysis
and pe cep ual colo simila i y me ics o achie e coa se wine
classi ica ion in uncons ained en i onmen s using sma phones.
Th ough machine lea ning, we iden i y colo cen oids o
ed, whi e, and os´
e wines, employing h eshold-based, Suppo
Vec o Machine (SVM), and closes -cen oid app oaches. By
le e aging he CIEDE2000 me ic alongside o he dis ance-based
echniques, ou SVM model achie es 98% classi ica ion accu acy
(F1-sco e = 0.979). This demons a es a cos -e ec i e, accessible
al e na i e o in asi e labo a o y me hods o wine colo analysis,
e en unde di e se, uncons ained condi ions.
Index Te ms—CIEDE2000, image p ocessing, uncons ained
en i onmen s, wine classi ica ion, wine colo analysis
I. INTRODUCTION
Wine colo con eys key in o ma ion abou quali y, o igin,
and aging, shaping consume expec a ions o a oma and as e
expe ience [1, 2]. Beyond aes he ics, wine colo e lec s i s
chemical composi ion, quali y, and p oduc ion p ocess, in lu-
encing bo h appea ance and senso y p o ile [3, 4].
T adi ional wine analysis me hods (senso y e alua ion,
spec opho ome y) a e accu a e ye expensi e and less acces-
sible o small-scale winemake s and consume s [1]. To add ess
hese challenges, digi al image analysis o e s a p omising
al e na i e, le e aging ad ancemen s in colo me ics and
compu a ional echniques [5]. The CIELAB colo space, ec-
ommended by he In e na ional O ganiza ion o Vine and
Wine (OIV), co ela es isually pe cei ed colo s wi h mea-
su ed pa ame e s [6]. Howe e , i s non-uni o mi ies led o he
CIEDE2000 o mula, imp o ing sub le colo di e en ia ion [5,
7]. In eg a ing digi al imaging wi h pe cep ual me ics like
CIEDE2000 enables wine colo analysis in di e se se ings,
bene i ing isually impai ed use s and e e yday consume s
who depend on accessible echnological solu ions o decision-
making.
To explo e a mo e lexible and accessible app oach o
wine colo analysis, we employed a combina ion o digi al
image p ocessing and colo di e ence me ics o classi y wines
based on hei isual cha ac e is ics, and explo e a lexible,
cos -e ec i e and accessible al e na i e ha can be applied
in di e se se ings. Digi al image-based echniques p o ide
a p omising app oach o pixel-le el colo analysis o ood
and be e age quali y, o e coming he limi a ions o adi ional
me hods [2, 4]. In ou wo k, wine images we e manually
segmen ed o isola e ele an a eas, ensu ing accu a e colo
assessmen [1]. The esul ing colo samples we e hen ana-
lyzed in bo h RGB and LAB colo spaces, which allowed o
a comp ehensi e unde s anding o wine colo cha ac e is ics.
Di e en colo di e ence me ics, including Euclidean,
Manha an, CIE94, and CIEDE2000 dis ance, ha e been used
o assess pe cep ual colo di e ences ela i e o clus e cen-
oids ob ained ia k-means clus e ing [7, 8]. The use o hese
me ics enables he iden i ica ion o ep esen a i e colo s o
each wine ype and minimizes he need o compa ison wi h
a b oad colo pale e. The CIEDE2000 me ic, in pa icula ,
p o ided a mo e consis en measu e o pe cei ed di e ences,
making i well-sui ed o dis inguishing among simila colo
ones in wine samples [5].
The p ima y con ibu ions o his s udy include:
•A da ase o wine images, segmen ed and anno a ed.
•The applica ion o k-means clus e ing in conjunc ion wi h
colo di e ence me ics o de ine ep esen a i e colo
cen oids o ed, whi e, and os´
e wines.
•A decision-making amewo k o wine classi ica ion
based on mul iple colo dis ance me ics.
Figu e 1 p esen s an o e iew o he p oposed pipeline o
wine pa ame e iza ion, om da a acquisi ion and egion-o -
in e es ex ac ion o colo e e ence iden i ica ion and h esh-
old de e mina ion.
Fig. 1. Pipeline illus a ing he o e all s eps o he wine pa ame e iza ion
p ocess using image analysis and colo me ics.
II. RELATED WORK
Colo simila i y measu emen is essen ial in digi al imaging,
pa icula ly o accu a ely cap u ing pe cep ual di e ences.
Howe e , adi ional Euclidean dis ance in RGB is limi ed
due o i s non-uni o mi y, while al e na i e me ics like Chi-
Squa e and Chebyshe Dis ance a e e ec i e in classi ica ion
and ou lie de ec ion. Likewise, adi ional Euclidean and non-
Euclidean me ics a e o en u ilized in colo spaces like
CIELAB o small colo di e ences. Fo la ge colo di e -
ences, hyb id me ics like HyAB, which combine Euclidean
and ci y-block me ics, o e enhanced e ec i eness by sepa-
a ely add essing hue, ch oma, and ligh ness [7]. The choice
o me ic depends on he applica ion and he na u e o he
colo da a being analyzed.
Pe cep ually uni o m colo spaces, no ably CIELAB, ha e
eme ged as he s anda d o pe cep ual colo analysis [5, 9].
The CIE76 o mula was ini ially in oduced bu had limi a ions
wi h high-ch oma colo s. To add ess his, CIE94 in oduced
weigh ing ac o s o ligh ness, ch oma, and hue, enhancing
pe cep ual di e en ia ion [9]. La e , CIEDE2000 imp o ed
u he upon CIE94 by ine- uning hese ac o s, and i has
since become he p e e ed s anda d in colo measu emen
[5]. Pa ha e e al. [10] and Mogol e al. [11] ha e highligh ed
how CIEDE2000 can be used o ood and be e age quali y,
whe e colo pe cep ion is c i ical. Mo e ecen ad ancemen s
ha e in oduced e icien CIEDE2000-based implemen a ions
in o de o educe compu a ional demands [8, 12].
In he con ex o wine, colo ac s as a key indica o o
quali y, o igin, aging, and au hen ici y, p ima ily in luenced
by phenolic compounds such as an hocyanins and annins.
B igh hues in young ed wines g adually da ken o e ime
due o an hocyanin deg ada ion and s able pigmen o ma ion
[13, 14]. T adi ional me hods, like he Glo ies me hod, assess
colo in ensi y and hue ia abso bance a speci ic wa eleng hs
[1]. Al hough spec opho ome ic echniques p o ide de ailed
and accu a e colo analysis, hey necessi a e cos ly equipmen ,
labo a o y condi ions, and specialized pe sonnel, hus limi ing
accessibili y [15].
To o e come hese cons ain s, digi al imaging has enabled
non-in asi e, cos -e ec i e me hods o wine colo analysis,
u ilizing colo spaces like RGB, HSI, and CIELAB o assess
au hen ici y and quali y while minimizing eliance on lab-
based echniques [16]. Digi al imaging has been success ully
employed o wine au hen ica ion, adul e a ion de ec ion, and
imp o ing consis ency in colo e alua ion h ough s anda d-
ized isual and pho og aphic echniques [17, 18]. While o -
e ing lexibili y and cos ad an ages, accu acy can be a ec ed
by ligh ing a iabili y and human pe cep ion.
III. DATASET ACQUISITION AND REPRESENTATION
A. Da ase P epa a ion
This s udy ocuses on ed, whi e, and os´
e wines due
o hei dis inc phenolic composi ions, quali y impac , and
consume ele ance [19]. Es ablished analy ical me hods o
hese wines a e well-documen ed, while me hodologies o
o he wine ypes emain less de eloped due o hei smalle
ma ke p esence [15].
Fo his s udy, we used a selec ion o wine images collec ed
h ough wo p ima y sou ces: (1) 75 images acqui ed using
a cus om-de eloped image acquisi ion applica ion, and (2)
a subse o 274 images selec ed om an online publicly
a ailable da ase [20]. Exis ing wine da ase s p ima ily ocus
on chemical analysis and mul ispec al imaging a he han
image-based classi ica ion. Examples include he UCI Wine
Da ase [21] and Vinho Ve de Da ase [22], which emphasize
chemical ea u es.
Each image was manually segmen ed o isola e he wine’s
egion o in e es (ROI) and minimize ex e nal in luences and
non-wine elemen s. Ex ac ed RGB colo alues we e s o ed
wi h co esponding wine labels, as RGB is he s anda d o ma
o consume de ices. Howe e , LAB, de i ed om RGB, is
mo e e ec i e o colo analysis by sepa a ing in ensi y om
ch oma ic in o ma ion, so bo h colo spaces we e used in ou
me hodology.
The low- esolu ion images, sou ced om he online da ase ,
we e upscaled using he supe - esolu ion Real-ESRGAN
(Real-Enhanced Supe -Resolu ion Gene a i e Ad e sa ial Ne -
wo k) model [23] o ensu e consis en egion-o -in e es (ROI)
ex ac ion. This is a deep lea ning-based model designed o
upscale images while p ese ing ine de ails and minimizing
a i ac s.
Be o e applying supe - esolu ion, each image was e alu-
a ed based on i s o al pixel coun . Images exceeding 1M
pixels we e deemed su icien ly de ailed and e ained wi hou
modi ica ion. Con e sely, images below his h eshold we e
upscaled using Real-ESRGAN, wi h a dynamically de e mined
scaling ac o : images smalle han 250kpixels (500 ×500)
we e upscaled by a ac o o 5, whe eas hose be ween 250k
and 1Mpixels we e upscaled by a ac o o 3.
Finally, he esul ing composi ion is shown in Table I, wi h
he inal esolu ion in e al o he en i e da ase anging om
a minimum o 1270 ×800 o a maximum o 3072 ×4096.
TABLE I
WINE COLOR CLASS COMPOSITION IN CUSTOM AND PUBLIC DATASET.
Red Whi e Ros´
e
Cus om Da ase 28 31 16
Public Da ase [20] 69 95 110
To al 97 126 126
B. Pixel Sampling S a egy and Da a Dis ibu ion
To ensu e a comp ehensi e and unbiased colo analysis, a
s uc u ed sampling me hodology is implemen ed. The p inci-
pal challenge s ems om he a iabili y in image dimensions,
which di ec ly impac s he numbe o a ailable pixels pe
sample. To mi iga e his issue, wo dis inc sampling s a egies
a e employed:
•Full-Sample Analysis: Random selec ion o a ixed num-
be o pixels, npixels, om all a ailable pixels wi hin an
image. By sampling ac oss he en i e image, his me hod
p o ides a holis ic ep esen a ion o he colo dis ibu ion
while main aining a con olled sample size. None heless,
a ia ions in image con en and illumina ion condi ions
may in oduce ce ain le els o bias.
•Cen al Region Sampling (Radius-Based Sampling):
To s anda dize he numbe o sampled pixels ac oss
di e en images, a ci cula egion cen e ed wi hin each
ROI is ex ac ed. Gi en a p ede ined numbe o pixels,
npixels, he sampling adius is compu ed as ollows:
=pnpixels ·Aimg/π
min(H, W)(1)
whe e Aimg ep esen s he o al image a ea, and Hand
Wdeno e he image heigh and wid h, espec i ely.
This echnique ensu es ha he sampled da a emains
consis en ac oss di e en images, he eby minimizing
po en ial sou ces o bias.
IV. COLOR CHARACTERIZATION AND CENTROID
ANALYSIS
A. Re e ence Colo s
In o de o iden i y ep esen a i e e e ence colo s o each
ype o wine, k-means clus e ing was used [24]. The goal o
his clus e ing app oach was o de e mine ep esen a i e colo s
o he wine ca ego ies unde s udy ( ed, whi e, and os´
e), hus
educing compu a ional complexi y.
The clus e ing analysis in ol ed mul iple expe imen s, a y-
ing sampling sizes, adius a ios, and me hods o calcula ing
cen al endencies. Table II summa izes he esul ing cen oid
coo dina es and clus e ing accu acy sco es, which e alua e
how well each sampling s a egy models he dis ibu ion o
wine colo s.
B. Cen oid Analysis
The s abili y and ep esen a i eness o cen oid coo dina es
we e e alua ed ac oss mul iple expe imen s o de e mine he
mos consis en clus e cen e s, as seen in Table II. Since
cen oids ac as e e ence poin s o colo classi ica ion, hei
eliabili y ac oss di e en sampling s a egies is essen ial o
ensu ing p ecise clus e ing and accu a e p edic ions.
•Red wine cen oid: The RGB cen oid o ed wine
emains highly s able ac oss expe imen s. Simila ly, LAB
cen oids exhibi minimal a ia ion, wi h L alues be-
ween 9 and 11 and only mino shi s in he A and
B channels. This consis ency indica es ha ed wine
clus e ing is less a ec ed by sampling a ia ions.
•Whi e and os´
e wine cen oid: Whi e and os´
e wine
cen oids display g ea e luc ua ion, pa icula ly in he
blue and g een channels, indica ing a b oade hue ange.
The LAB coo dina es con i m his a iabili y, wi h L
alues o whi e wine anging om 70 o 71 and o os´
e
wine om 51 o 52. These indings sugges ha hese
ca ego ies a e mo e sensi i e o sampling di e ences.
As shown in Table II, Expe imen 3 deli e s he mos
s able and accu a e cen oids o all h ee wine ypes. These
coo dina es align wi h he highes classi ica ion accu acies (up
o 0.7542 in RGB and 0.8966 in LAB), making hem he bes
ep esen a i es o each colo ca ego y.
Figu e 2 compa es k-means clus e ing esul s in LAB and
RGB spaces. LAB clus e ing achie es be e sepa a ion, es-
pecially o ed wine, aligning closely wi h ue labels. In
con as , RGB clus e ing shows mo e o e lap be ween os´
e
and whi e wines, indica ing LAB p o ides a mo e dis inc
ep esen a ion o classi ica ion.
Fig. 2. Example o K-Means clus e ing esul s o LAB and RGB sampled
da a, using 2,000 pixel samples pe image in a adius o 0.5. This compa ison
illus a es he p edic ed clus e colo s e sus he ue labels.
V. COLOR-BASED WINE CLASSIFICATION
This s age ocuses on classi ying wine samples based on
hei colo a ibu es by sys ema ically ex ac ing key ea u es
om digi al images, compu ing ad anced dis ance me ics o
cen oids, and classi y he sample.
A. Decision F amewo k o Wine Classi ica ion
Wine classi ica ion equi es ansi ioning om pixel-le el
analysis, whe e each pixel is assessed indi idually, o an agg e-
ga ed decision ha e alua es he en i e egion o in e es .Fo
each ex ac ed ROI, all pixels a e compa ed agains e e ence
cen oids using dis ance me ics. The classi ica ion is hen
de e mined based on he h ee ollowing s a egies:
•Closes -Cen oid Classi ica ion: The median dis ance o
all pixels o each cen oid is compu ed, and he sample
is classi ied in o he wine ype wi h he smalles median
dis ance.
•Th eshold-Based Classi ica ion: A wine sample is clas-
si ied on a pixel-by-pixel basis, whe e each pixel’s dis-
ance o p ede ined cen oids is measu ed and h esh-
olded. Consequen ly, he inal classi ica ion o he wine
image is de e mined by majo i y o e.
•SVM-based Classi ica ion: Each wine sample is classi-
ied a he pixel le el. Fo each pixel, i s dis ances o he
e e ence cen oids a e compu ed using a speci ic colo -
based me ic. These h ee dis ances se e as inpu ea u es
o an SVM classi ie , which de e mines he inal wine
classi ica ion based on majo i y o ing ac oss all pixels
in an image.
TABLE II
CENTROID COORDINATES AND ACCURACY SCORES FOR DIFFERENT SAMPLING STRATEGIES IN WINE COLOR CLUSTERING.
Expe imen ID Radius Ra io Numbe o Pixels RGB Cen oid Coo dina es LAB Cen oid Coo dina es RGB Accu acy Sco e LAB Accu acy Sco e
Red Whi e Ros´
e Red Whi e Ros´
e
1 Full Image 5000 (39, 22, 18) (220, 192, 141) (178, 100, 53) (11, 6, 4) (71, 0, 31) (51, 39, 42) 0.7262 0.8606
2 0.25 2000 (37, 22, 18) (218, 190, 131) (183, 97, 46) (10, 5, 4) (70, -1, 33) (52, 41, 45) 0.7498 0.8932
3 0.50 2000 (35, 20, 17) (218, 188, 129) (181, 97, 46) (9, 5, 4) (70, 0, 34) (52, 40, 45) 0.7542 0.8966
4 0.75 2000 (35, 20, 17) (219, 190, 132) (181, 97, 47) (9, 5, 4) (70, 0, 33) (52, 40, 44) 0.7533 0.8918
B. Dis ance Me ics o Colo Compa ison
To classi y wine samples, hei colo p o iles a e compa ed
wi h cen oids ep esen ing ed, whi e, and os´
e wines in
bo h RGB and LAB colo spaces. The ollowing me ics
a e widely employed o quan i y colo simila i y, p o iding
dis inc app oaches o assess bo h pe cep ual and nume ical
di e ences in colo .
Euclidean Dis ance: The Euclidean dis ance quan i ies he
di ec geome ic sepa a ion be ween wo poin s in a gi en colo
space. In he LAB space, his measu e co esponds o ∆E76.
Manha an Dis ance: The Manha an dis ance compu es
he o al absolu e di e ences be ween co esponding compo-
nen s o wo colo ec o s. I emphasizes linea de ia ions
along each dimension.
CIE94 Dis ance (∆E94): Expanding upon ∆E76, he
CIE94 me ic inco po a es pe cep ual weigh ing ac o s ha
accoun o a ia ions in ligh ness, ch oma, and hue. These
adjus men s imp o e he alignmen be ween nume ical colo
di e ences and human isual pe cep ion [25].
CIEDE2000 Dis ance (∆E00): Re ines ∆E94 by in oduc-
ing addi ional pe cep ual co ec ions, including ch oma and
hue in e ac ions. These e inemen s enhance he accu acy o
colo di e ence assessmen s [8].
Fo each dis ance me ic, he dis ances be ween all sample
pixels and he p ede ined cen oids o he espec i e a e
compu ed. The inal image classi ica ion is achie ed based on
majo i y o ing ac oss all pixels in an image. This p ocess is
independen ly pe o med o each dis ance me ic, acili a ing
mul iple assessmen s o colo simila i y.
C. Th eshold de e mina ion
Selec ing an app op ia e classi ica ion h eshold accoun s
o na u al colo a iabili y in each wine ca ego y. Th esholds
we e de e mined by analyzing he dis ance dis ibu ions using
s a is ical me ics, wi h he median chosen as he mos obus
app oach o minimize ou lie s. Table III p esen s he inal
h esholds o each ca ego y.
Figu e 3 depic s he di e en CIEDE2000 dis ance dis ibu-
ions o andom sampling using he e e ence cen oids. Each
his og am highligh s he mean and median alues, guiding
h eshold de e mina ion o accu a e classi ica ion.
The selec ion p ocess ensu es ha cen oids ep esen each
wine ca ego y, whe eas h esholds classi y wines based on
pe cep ually signi ican di e ences.
VI. RESULTS & DISCUSSION
Table IV p o ides an o e iew o he p ecision, ecall,
F1-sco e, and accu acy achie ed by h ee dis inc classi-
ica ion s a egies—Closes -Cen oid, Th eshold-Based, and
SVM-based—ac oss bo h RGB and LAB colo spaces. Wi hin
he RGB colo space, we explo e Euclidean and Manha an
dis ance measu es. In he LAB colo space, we examine
Manha an LAB, CIE76, CIE94, and CIEDE2000.
The Closes -Cen oid me hod ends o achie e lowe pe -
o mance in he RGB colo space. The ac ha he RGB
space may no accu a ely ep esen pe cep ual colo dispa i ies.
Speci ically, he Euclidean and Manha an RGB esul s a ely
su pass 0.35 in F1-sco e, unde sco ing he inhe en limi a ions
o pu ely dis ance-based classi ica ion wi hin a non-pe cep ual
colo space. When ansi ioning o LAB, his same app oach
shows subs an ial imp o emen , pa icula ly unde s anda d
pe cep ual me ics like CIE76, CIE94, and CIEDE2000.
Th eshold-Based classi ica ion displays mixed pe o mance.
In he RGB space, i pe o ms sligh ly be e han Closes -
Cen oid, indica ing ha well-de ined dis ance h esholds may
cap u e sub le colo a ia ions mo e e ec i ely. Howe e , he
gap be ween p ecision and ecall can s ill be conside able.
Se ing na owe h esholds migh ampli y de ec ion o ce ain
wine ypes bu isks o e looking o he s. Likewise, using
pe cep ual me ics in LAB subs an ially aises o e all pe -
o mance and ends o balance p ecision and ecall, al hough
conse a i e h esholds some imes lead o high p ecision a
he expense o ecall.
The SVM-based app oach consis en ly demons a es supe-
io pe o mance ac oss nea ly all es ed scena ios. I achie ed
F1-sco es abo e 0.90 in RGB and app oaching o exceeding
0.95 in LAB wi h pe cep ual me ics, adap ing mo e lexibly
o complex decision bounda ies. This ad an age becomes
mo e p onounced when using LAB me ics, whe e pe cep ual
di e ences among ed, whi e, and os´
e a e cap u ed mo e
e ec i ely. No ably, unde CIEDE2000, he SVM-based clas-
si ie a ains bes p ecision, ecall, and F1-sco es.
The con ibu ions o his s udy ex end beyond classi i-
ca ion accu acy, encompassing aspec s o p ac icali y and
implemen a ion. An impo an con ibu ion is he in eg a ion
o sma phones o bo h da a acquisi ion and p ocessing.
In his con ex , he sma phone unc ions no only as an
imaging de ice bu also as a compu a ional pla o m, enabling
a po able, cos -e ec i e, and eal- ime solu ion o wine
classi ica ion. As men ioned, his se up allows ope a ion in
uncons ained en i onmen s, wi hou he need o specialized
labo a o y equipmen . The use o eadily a ailable consume -
g ade sma phones u he enhances accessibili y, making he
app oach iable o end-use s and small-scale p oduce s. The
es s using a Samsung Galaxy A40 (Exynos 7904, 4 GB RAM)
Fig. 3. Dis ibu ions o dis ances om each cen oid o h eshold de e mina ion.
TABLE III
THRESHOLD VALUES FOR WINE CLASSIFICATION USING DIFFERENT DISTANCE METRICS.
Wine Type Euclidean (RGB) Manha an (RGB) CIE76 Manha an (LAB) CIE94 CIEDE2000
Red 25.4 39.0 9.3 14.9 8.7 7.8
Whi e 68.95 111.0 19.29 27.29 14.89 12.0
Ros´
e60.0 91.0 18.35 28.35 11.33 11.14
TABLE IV
PRECISION,RECALL, F1-SCORE,AND ACCURACY FOR EACH COMBINATION OF CLASSIFICATION APPROACH AND METRIC IN RGB AND LAB SPACES.
Classi ica ion App oach Euclidean RGB Manha an RGB Manha an LAB
P ecision Recall F1-Sco e Accu acy P ecision Recall F1-Sco e Accu acy P ecision Recall F1-Sco e Accu acy
Closes -Cen oid 0.308 0.300 0.301 0.300 0.222 0.217 0.179 0.217 0.913 0.905 0.904 0.905
Th eshold-Based 0.311 0.306 0.306 0.306 0.356 0.309 0.326 0.309 0.968 0.816 0.883 0.816
SVM-based 0.929 0.931 0.930 0.928 0.871 0.876 0.872 0.868 0.951 0.953 0.952 0.951
CIE76 CIE94 CIE2000
P ecision Recall F1-Sco e Accu acy P ecision Recall F1-Sco e Accu acy P ecision Recall F1-Sco e Accu acy
Closes -Cen oid 0.915 0.908 0.908 0.908 0.879 0.873 0.872 0.873 0.903 0.896 0.896 0.896
Th eshold-Based 0.930 0.790 0.850 0.790 0.987 0.796 0.880 0.796 0.974 0.808 0.880 0.808
SVM-based 0.947 0.951 0.949 0.948 0.976 0.977 0.977 0.977 0.979 0.980 0.979 0.980
demons a ed ha he comple e pipeline including image cap-
u e, ROI ex ac ion, pixel sampling, and classi ica ion using
an SVM model wi h he CIEDE2000 me ic can be pe o med
in an a e age o 50 milliseconds pe image.
These expe imen s highligh how inco po a ing pe cep ual
colo ep esen a ions and decision-making me hods can no-
ably imp o e classi ica ion accu acy. By using colo me ics
and machine lea ning echniques, he app oaches es ed we e
able o cap u e dis inc ions among ed, whi e, and os´
e wines.
VII. CONCLUSION
This s udy p esen s an image-based me hod o coa se-
g ained wine colo classi ica ion using LAB colo space
and he CIEDE2000 me ic. Ou pe o ming RGB me hods,
e alua ions on ed, whi e, and os´
e wines show ha using
op imized colo cen oids wi h CIEDE2000 achie es high
accu acy despi e ligh ing and acquisi ion challenges.
The p oposed app oach enables eal- ime wine colo anal-
ysis o consume s, isually impai ed use s, and po en ial
mul imodal in eg a ion wi h addi ional sensing echnologies.
Fu u e esea ch may ex end he scope o wine ca ego ies
analyzed and in eg a e ad anced calib a ion echniques o
u he a enua e en i onmen al in luences. Addi ionally, ex-
panding he da ase —po en ially h ough c owdsou ced image
con ibu ions—could enhance he model’s obus ness agains
g ea e in a-class colo a iabili y.
ACKNOWLEDGMENT
This wo k was conduc ed as pa o he Wa son p ojec ,
unded by he Eu opean Union’s Ho izon Eu ope esea ch and
inno a ion p og amme, unde g an ag eemen No. 101084265.
DISCLAIMER
Funded by he Eu opean Union. Views and opinions ex-
p essed a e howe e hose o he au ho (s) only and do no
necessa ily e lec hose o he Eu opean Union o he Eu o-
pean Resea ch Execu i e Agency (REA). Nei he he Eu opean
Union no he g an ing au ho i y can be held esponsible o
hem.
REFERENCES
[1] Ma cel Hensel e al. “New Insigh s in o Wine Colo
Analysis: A Compa ison o Analy ical Me hods o Sen-
so y Pe cep ion o Red and Whi e Va ie al Wines”. In:
Jou nal o Ag icul u al and Food Chemis y 72 (2024),
pp. 2008–2017.
[2] Di Wu and Da-Wen Sun. “Colou measu emen s by
compu e ision o ood quali y con ol: A e iew”.
In: T ends in Food Science & Technology 29 (2013),
pp. 5–20.
[3] Ma ia Lou des Gonzalez-Mi e e al. “Measu ing colou
appea ance o ed wines”. In: Food Quali y and P e e -
ence 18 (2007), pp. 862–871.
[4] Ca los Mon ei o de Lima e al. “Digi al image-based
acing o geog aphic o igin, winemake , and g ape ype
o ed wine au hen ica ion”. In: Food Chemis y 312
(2020), p. 126060.
[5] Gau a Sha ma, Wencheng Wu, and Edul N. Dalal.
“The CIEDE2000 colo -di e ence o mula: Implemen-
a ion no es, supplemen a y es da a, and ma hema ical
obse a ions”. In: Colo Resea ch & Applica ion 30.1
(2005), pp. 21–30. DOI: h ps://doi.o g/10.1002/col.
20070. URL: h ps://onlinelib a y.wiley.com/doi/abs/10.
1002/col.20070.
[6] Dominik Du ne and Ma cel Hensel. “Is he e a need o
e-de ine he me hods o e alua e wine colo ”. In: BIO
Web o Con e ences 68 (2023), p. 02008.
[7] Saeedeh Abasi, Mohammad Amani Teh an, and Ma k D
Fai child. “Dis ance me ics o e y la ge colo di e -
ences”. In: Colo Resea ch & Applica ion 30 (2019),
pp. 1–16.
[8] Am´
e ico Pe ei a e al. “E icien CIEDE2000-Based
Colo Simila i y Decision o Compu e Vision”. In:
IEEE T ansac ions on Ci cui s and Sys ems o Video
Technology 30.7 (2020), pp. 2141–2154. DOI: 10.1109/
TCSVT.2019.2914969.
[9] Ming Luo, Guihua Cui, and B. Rigg. “The de el-
opmen o he CIE 2000 colou -di e ence o mula:
CIEDE2000”. In: Colo Resea ch & Applica ion 26
(Oc . 2001), pp. 340–350. DOI: 10.1002/col.1049.
[10] Pankaj B Pa ha e, Umezu uike Linus Opa a, and Fahad
Al-Julanda Al-Said. “Colou measu emen and analysis
in esh and p ocessed oods: A e iew”. In: Food and
biop ocess echnology 6 (2013), pp. 36–60.
[11] Bu c¸e A ac¸ Mogol and Vu al G¨
okmen. “Compu e
ision-based analysis o oods: A non-des uc i e colou
measu emen ool o moni o quali y and sa e y”. In:
Jou nal o he science o ood and ag icul u e 94 (May
2014). DOI: 10.1002/js a.6500.
[12] Am´
e ico Pe ei a, Ped o Ca alho, and Luis Co e-
Real. “Boos ing colo simila i y decisions using he
CIEDE2000 PF Me ic”. In: Signal, Image and Video
P ocessing 16.7 (2022), pp. 1877–1884.
[13] Kassa a S ella, Pa ke Mango, e al. “Phenolic Compo-
si ions o 50 and 30 Yea Sequences o Aus alian Red
Wines: The Impac o Wine Age”. In: (2012).
[14] Ma ´
ıa-Pila S´
aenz-Na ajas e al. “Pigmen composi ion
and colo pa ame e s o comme cial Spanish ed wine
samples: Linkage o quali y pe cep ion”. In: Eu opean
Food Resea ch and Technology 232 (2011), pp. 877–
887.
[15] Ma cel Hensel e al. “Speci ica ion and Simpli ica ion
o Analy ical Me hods o De e mine Wine Colo ”. In:
P ocesses 10.12 (2022). ISSN: 2227-9717. DOI: 10 .
3390/p 10122707. URL: h ps://www.mdpi.com/2227-
9717/10/12/2707.
[16] AN Unde hill, CD Hi sch, and MD Cla k. “E alua ing
and mapping g ape colo using image-based pheno yp-
ing”. In: Plan Phenomics (2020).
[17] Ca los He e o-La o e e al. “De ec ion and quan i i-
ca ion o adul e a ions in aged wine using RGB digi al
images combined wi h mul i a ia e chemome ic ech-
niques”. In: Food chemis y: X 3 (2019), p. 100046.
[18] Ma k D. Fai child. “Visual and pho og aphic assessmen
o wine colo ”. In: Colo Resea ch & Applica ion 48.1
(2023), pp. 21–31. DOI: h ps://doi.o g/10.1002/col.
22787. ep in : h ps://onlinelib a y.wiley.com/doi/pd /
10.1002/col.22787. URL: h ps://onlinelib a y.wiley.
com/doi/abs/10.1002/col.22787.
[19] BANC Roxana e al. “Phenolic composi ion and an iox-
idan ac i i y o ed, os´
e and whi e wines o igina ing
om Romanian g ape cul i a s”. In: No ulae Bo anicae
Ho i Ag obo anici Cluj-Napoca 48.2 (2020), pp. 716–
734.
[20] P. G. Miguel de Oli ei a. Wine Glass Da ase . Re ie ed
om Kaggle. 2022. URL: h ps : / / www. kaggle . com /
da ase s/mig ande lei/ es eda ase wine.
[21] S e an Aebe ha d and M. Fo ina. Wine.
UCI Machine Lea ning Reposi o y. DOI:
h ps://doi.o g/10.24432/C5PC7J. 1992.
[22] Paulo Co ez e al. “Modeling wine p e e ences by da a
mining om physicochemical p ope ies”. In: Decision
Suppo Sys ems 47.4 (2009). Sma Business Ne wo ks:
Concep s and Empi ical E idence, pp. 547–553. ISSN:
0167-9236. DOI: h ps://doi.o g/10.1016/j.dss.2009.
05.016. URL: h ps://www.sciencedi ec .com/science/
a icle/pii/S0167923609001377.
[23] Xin ao Wang e al. Real-ESRGAN: T aining Real-Wo ld
Blind Supe -Resolu ion wi h Pu e Syn he ic Da a. 2021.
a Xi : 2107.10833 [eess.IV].URL: h ps://a xi .o g/
abs/2107.10833.
[24] Ch is ophe M Bishop. “Pa e n ecogni ion and ma-
chine lea ning”. In: Sp inge google schola 2 (2006),
pp. 1122–1128.
[25] Rode ick McDonald and Kenne h J Smi h. “CIE94-a
new colou -di e ence o mula”. In: Jou nal o he Soci-
e y o Dye s and Colou is s 111.12 (1995), pp. 376–379.
