DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
Ensemble Fea u e Selec ion App oach Based on
Fea u e Ranking o Rice Seed Images
Classi ica ion
Dzi Lam T an TUAN 1, Thongchai SURINWARANGKOON 2, Ki ikhun MEETHONGJAN 3,
Vinh T uong HOANG1
1Depa men o Image P ocessing and Compu e G aphics, Facul y o Compu e Science,
Ho Chi Minh Ci y Open Uni e si y, 97 Vo Van Tan S ee , 700000 Ho Chi Minh Ci y, Vie nam
2Depa men o Business Compu e , Facul y o Managemen Science,
Suan Sunandha Rajabha Uni e si y, 1 U Thong Nok Rd, 10300 Bangkok, Thailand
3Depa men o Applied Science, Facul y o Science and Technology,
Suan Sunandha Rajabha Uni e si y, 1 U Thong Nok Rd, 10300 Bangkok, Thailand
[email p o ec ed], hongc[email p o ec ed], ki ikh[email p o ec ed], [email p o ec ed]
DOI: 10.15598/aeee. 18i3.3726
Abs ac . In sma ag icul u e, ice a ie y inspec ion
sys ems based on compu e ision need o be used o
ecognizing ice seeds ins ead o using echnical expe s.
In his pape , we ha e in es iga ed h ee ypes o lo-
cal desc ip o s, such as Local Bina y Pa e n (LBP),
His og am o O ien ed G adien s (HOG) and GIST
o cha ac e ize ice seed images. Howe e , his ap-
p oach aises he cu se o dimensionali y phenomenon
and needs o selec he ele an ea u es o a compac
and be e ep esen a ion model. A new ensemble ea-
u e selec ion is p oposed o ep esen all use ul in o -
ma ion collec ed om di e en single ea u e selec ion
me hods. The expe imen al esul s ha e shown he e -
iciency o ou p oposed me hod in e ms o accu acy.
Keywo ds
Ensemble ea u e selec ion, ea u e anking,
ea u e selec ion, GIST, HOG, LBP, ice seed
image.
1. In oduc ion
Rice is he mos impo an ood sou ce o people in
many coun ies including Asia, A ica, La in Ame -
ica, and he Middle Eas . P oduc s made om ice,
including ice p oduc s and indi ec p oduc s, a e in-
dispensable in he daily meals o billions o people
a ound he wo ld. Nowadays, mo e ice a ie ies a e
c ea ed wi h di e si ied quali y and p oduc i i y. Di -
e en a ie ies o ice can be mixed du ing cul i a ion
and ading. We p ac ically need o de elop a sys em
o au oma ically iden i y ice seeds based on machine
ision. Va ious wo ks ha e been p oposed o au o-
ma ic inspec ion and quali y con ol in ag icul u e [10].
In he pas decade, a g ea numbe o local image de-
sc ip o s [13] ha e been p oposed o cha ac e izing im-
ages. Each kind o a ibu e ep esen s he da a in
a speci ic space and has p ecise spa ial meaning and
s a is ical p ope ies.
Di e en local desc ip o s a e ex ac ed o c e-
a e a mul i- iew image ep esen a ion, like Local Bi-
na y Pa e n (LBP), His og am o O ien ed G adi-
en s (HOG), and GIST. Nha and Hoang [20] p esen
a me hod o use he ea u es ex ac ed om h ee
desc ip o s (Local Bina y Pa e n, His og am o O i-
en ed G adien and GIST) o acial images classi-
ica ion. The conca ena ed ea u es a e hen ap-
plied by canonical co ela ion analysis o ha e a com-
pac ep esen a ion be o e eeding in o he classi ie .
Van and Hoang [24] p opose o educe noisy and i -
ele an Local Te na y Pa e n (LTP) ea u es and
HOG coding on di e en colo spaces o ace anal-
ysis. Hoai e al. [12] in oduce a compa a i e s udy
o hand-c a ed desc ip o s and Con olu ional Neu-
al Ne wo ks (CNN) o ice seed images classi ica-
ion. Meba sion e al. [17] use Fou ie desc ip-
o s and h ee geome ical ea u es o ce eal g ains
ecogni ion. Duong and Hoang [9] apply o ex ac
ice seed images based on ea u es coded in mul i-
ple colo spaces using HOG desc ip o . Mul i- iew
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 198
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
lea ning was in oduced o complemen in o ma ion
be ween di e en iews. While conca ena ing di e -
en ea u e se s, i is e iden ha all he ea u es do
no gi e he same con ibu ion o he lea ning ask
and some ea u es migh dec ease he pe o mance.
Thus, ea u e selec ion me hods a e applied as a p e-
p ocessing s age o high-dimensional ea u e space.
I in ol es selec ing pe inen and use ul ea u es, while
a oiding and igno ing edundan and i ele an in o -
ma ion [26]. A no el eache -s uden ea u e selec ion
app oach [19] is p oposed o ind he bes ep esen a-
ion o da a in low dimension.
Recen ly, ensemble ea u e selec ion has eme ged as
a new app oach ha p omises o enhance he obus -
ness and pe o mance. I is he p ocess o pe o ming
di e en ea u e selec ion in o de o ind an op imum
subse o ea u es. Ins ead o using a single selec ion
app oach, an ensemble me hod combines he esul s
o di e en app oaches in o a inal single subse o ea-
u es. Seijo-pa do e al. [23] p opose o combine di -
e en ea u e selec ion app oaches on he e ogeneous
da a based on a p ede ined h eshold alue. Chiew e
al. [5] in oduce a hyb id ensemble ea u e selec ion
based on Cumula i e Dis ibu ion Func ion g adien .
This me hod can de e mine an es ima ion o ea u e
cu -o au oma ically. D o a e al. [8] p opose a new
ensemble ea u e selec ion app oach me hods based on
di e en o ing echniques such as plu ali y, and Bo da
coun . A comple e and de ailed e iew o ensemble ea-
u e selec ion me hods is in oduced in [3].
In his pape , we p opose a new ensemble ea u e se-
lec ion app oach based on mul i- iew desc ip o s (LBP,
HOG and GIST) ex ac ed om ice seed images. Se -
e al ea u e selec ion app oaches a e u he in es i-
ga ed and combined o ind an op imum subse o ea-
u es wi h he pu pose o enhance he classi ica ion
pe o mance. This pape is o ganized and s uc u ed
as ollows. Sec ion 2. , in oduces he ea u e ex ac -
ing me hods based on h ee local image desc ip o s.
Sec ion 3. p esen s a p oposed ensemble ea u e
selec ion amewo k. Sec ion 4. shows expe imen-
al esul s. Finally, he conclusion is hen p o ided in
Sec. 5. .
2. The Fea u e Ex ac ing
Me hods
This sec ion b ie ly e iews h ee local image desc ip-
o s used in expe imen s o ea u e ex ac ion.
2.1. Local Bina y Pa e n
The LBPP,R(xc, yc)code o each pixel (xc, yc)is cal-
cula ed by compa ing he g ay alue gco he cen al
pixel wi h he g ay alues {gi}P−1
i=0 o i s Pneighbo s,
as ollows [21]:
LBPP,R =
P−1
X
p=0
ω(gp−gc)2p,(1)
whe e gcis he g ay alue o cen al, gpis he g ay
alue o P,Ris he adius o he ci cle, and ω(gp−gc)
is de ined as:
ω(gp−gc) = 1i (gp−gc)≥0,
0o he wise.(2)
2.2. GIST
GIST is i s ly p oposed by Oli a and To alba [22]
in o de o classi y objec s which ep esen he shape
o he objec . The p ima y idea o his app oach is
based on he Gabo il e :
h(x, y) = e
−1
2x2
δ2
x
+y2
δ2
ye−j2π(u0x+ 0y).(3)
Fo each (δx,δy) o he image ia he Gabo il-
e , we ob ain all he image elemen s ha a e close
o he poin colo (u0x+ 0y). The esul o he cal-
cula ed Vec o GIST will ha e many dimensions.
To educe he size o he ec o , we a e aged each 4×4
g id o he abo e esul s. Each image also con igu es
a Gabo il e wi h 4 scales and 8 di ec ions (o ien-
a ions), c ea ing 32 cha ac e is ic maps o he same
size.
2.3. His og ams o O ien ed
G adien
HOG desc ip o s a e applied o di e en asks in ma-
chine ision [7] such as human de ec ion [6]. HOG ea-
u e is ex ac ed by coun ing he occu ences o g adi-
en o ien a ion based on he g adien angle and he g a-
dien magni ude o local pa ches o an image. The g a-
dien angle and magni ude a each pixel is compu ed
in an 8×8pixels pa ch. Nex , 64 g adien ea u e ec-
o s a e di ided in o 9 angula bins 0–180◦(20◦each).
The g adien magni ude Tand angle Ka each posi-
ion (k,h) om an image Ja e compu ed as ollows:
∆k=|J(k−1, h)−J(k+ 1, h)|.(4)
∆h=|J(k, h −1) −J(k, h + 1)|.(5)
T(k, h) = q∆2
i+ ∆2
j.(6)
K(k, h) = an−1∆k
∆j.(7)
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 199
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
Da a
( ea u ed desc ip ion)
T aining
se
Tes ing
se
Fea u e
Selec ions
Final
model
Acc
Top 2
spli op
Ranking
Acc
Top 3
spli op
Ranking
Acc
Top 1
combine
New anking
spli op
Ranking combine
Fea u e
Selec ion
Final
model Acc
Rankings
las
Ranking
Fig. 1: The p oposed ensemble ea u e selec ion app oach.
3. Ensemble Fea u e Selec ion
The dimension educ ion has se e al ad an ages
and impac s on da a s o age, gene aliza ion capabil-
i y and compu ing ime. Based on he a ailabili y
o supe ised in o ma ion (i.e, class labels), ea u e
selec ion echniques can be g ouped in o wo la ge
ca ego ies: supe ised and unsupe ised con ex [1].
Addi ionally, di e en s a egies o ea u e selec ion a e
p oposed based on e alua ion p ocesses such as il e ,
w appe and hyb id me hods [11]. Hyb id app oaches
inco po a e bo h il e and w appe in o a single s uc-
u e, in o de o gi e an e ec i e solu ion o dimen-
sionali y educ ion [4]. In o de o s udy he con ibu-
ion o ea u e selec ion app oaches o ice seed images
classi ica ion, we p opose o apply se e al selec ion ap-
p oaches based on images ep esen ed by mul i- iew
desc ip o s. In he ollowing subsec ion, we will sho ly
p esen he common ea u e selec ion me hods applied
in supe ised lea ning con ex .
•LASSO (Leas Absolu e Sh inkage and Selec ion
Ope a o ) allows o compu e ea u e selec ion
based on he assump ion o linea dependency be-
ween inpu ea u es and ou pu alues. Lasso
minimizes he sum o squa es o esiduals when
he sum o he absolu e alues o he eg es-
sion coe icien s is less han a cons an , which
yields ce ain s ic eg ession coe icien s equal
o 0 [4] and [25].
•mRMR (Maximum Rele ance and Minimum Re-
dundancy) is a mu ual in o ma ion based ea u e
selec ion c i e ion, o dis ance/simila i y sco es
o selec ea u es. The aim is o penalize a ea-
u e’s ele ance by i s edundancy in he p esence
o he o he selec ed ea u es [27].
•Relie F [15] is ex ended om Relie [14] o suppo
mul iclass p oblems. Relie F seems o be a p omis-
ing heu is ic unc ion ha may o e come he my-
opia o cu en induc i e lea ning algo i hms. Ki a
and Rendell used Relie F as a p ep ocesso o elim-
ina e i ele an a ibu es om da a desc ip ion
be o e lea ning. Relie F is gene al, ela i ely e i-
cien , and eliable enough o guide he sea ch in
he lea ning p ocess [16].
•CFS (Co ela ion Fea u e Selec ion) mainly ap-
plies heu is ic me hods o e alua e he e ec
o a single ea u e co esponding o each g oup in
o de o ob ain he op imal subse o a ibu es.
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 200
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
•Fishe [2] iden i ies a subse o ea u es so ha
he dis ances be ween samples in di e en classes
a e as la ge as possible, while he dis ances be-
ween samples in he same class a e as small as
possible. Fishe selec s he op anked ea u es
acco ding o i s sco es.
•ILFS (In ini e La en Fea u e Selec ion) is a ech-
nique consis s o h ee s eps such as p ep ocessing,
ea u e weigh ing based on a ully connec ed g aph
in each node ha connec all ea u es. Finally, en-
e gy sco es o he pa h leng h a e calcula ed, hen
ank i s co espondence wi h he ea u e [18].
Figu e 1 p esen he p oposed ensemble ea u e selec-
ion amewo k. Each indi idual ea u e selec ion ap-
p oach has i s p os and cons, he aim o his p oposi ion
is o combine he p os o di e en me hods o boos
he pe o mance in e ms o accu acy. We p opose
o apply h ee independen ea u e selec ion me hods
o selec he "bes " subse o ea u es. Then, a new
anking me hod is applied o combined ea u e space.
This can inc ease he dimension space bu i allows
o collec ele an ea u es de e mined by di e en se-
lec ion me hods. The meaning behind is o selec
he mos ele an ea u es so ha we ha e o apply
a inal anking o elimina e he edundan and noisy
ea u es.
4. Expe imen al Resul s
4.1. Expe imen al Se up
(a) BC-15. (b) Nep-87.
(c) Huong hom 1. (d) Thien uu 8.
(e) Q-5. ( ) Xi-23.
Fig. 2: The p oposed o ensemble ea u e selec ion.
The ice seed images da abase comp ises six ice
seed a ie ies in he no he n Vie nam (illus a ed in
Fig. 2) [9]. We apply he 1-NN and SVM classi ie s
o e alua e he classi ica ion pe o mance ia accu acy
a e. A hal o he da abase is selec ed o he aining
se and he es is used o he es ing se . We use
Hold-ou me hod wi h a io (1/2 and 1/2) and spli
he aining and es ing se by chessboa d decomposi-
ion. All expe imen s a e implemen ed and simula ed
by Ma lab 2019a and conduc ed on a PC wi h a con-
igu a ion o a CPU Xeon 3.08 GHz, 64 GBs o RAM.
4.2. Resul s
Table 1 shows he accu acy ob ained by 1-NN and
SVM classi ie when no ea u e selec ion app oach is
applied. The i s column indica es he ea u es used
o ep esen ing images. We use h ee indi idual local
desc ip o s namely LBP, GIST, and HOG and he con-
ca ena ion o "LBP + GIST" ea u es. The second
column indica es he numbe o ea u es (o dimen-
sion) co esponding o ea u es ype. The hi d and
ou h columns show he accu acy ob ained by 1-NN
and SVM classi ie . We obse e ha he mul i- iew
by conca ena ing mul iple ea u es gi es be e pe -
o mance, howe e i inc eases he dimension. Hence,
he pe o mance o SVM classi ie is be e han 1-NN
classi ie wi h 94.7 % o accu acy.
Tab. 1: Classi ica ion pe o mance wi hou selec ion app oach
o di e en ypes o ea u es.
Fea u es Dimension 1-NN SVM
LBP 768 53.0 77.0
GIST 512 69.4 88.3
HOG 21.384 71.5 94.7
LBP + GIST 1.280 70.5 91.7
The ollowing ables and igu es illus a e in de ailed
he classi ica ion in single o mul i- iew based on h ee
desc ip o s:
•LBP: Table 2, Fig. 3(a) and Fig. 3(b),
•GIST: Table 4, Fig. 4(a) and Fig. 4(b),
•HOG: Table 5, Fig. 5(a) and Fig. 5(b),
•LBP + GIST: Table 3, Fig. 6(a) and Fig. 6(b).
Table 2 and Fig. 3 show ha he classi ica ion pe -
o mance each 53.0 % by 1-NN classi ie on LBP de-
sc ip o . A e using 6 di e en ea u e selec ion ap-
p oaches, we ob ain h ee bes candida es wi h de-
scendan accu acy such as mRMR (59.0 %), ILFS
(58.4 %) and Relie F (54.2 %). Based on he p o-
posed me hod illus a ed in Fig. 1, he 85 % pe cen age
o selec ed ea u es by Relie F is combined wi h 43 %
o selec ed ea u e de e mined by ILFS me hod. We
ob ain he new subse o ea u es which is calcula ed
as ollows:
(768 ·0.85) + (768 ·0.43) = 983 dim.(8)
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 201
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
20
25
30
35
40
45
50
55
60
Accu acy
1-NN classi ie on LBP ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(a) 1-NN.
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
30
40
50
60
70
80
90
Accu acy
SVM classi ie on LBP ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(b) SVM.
Fig. 3: 1-NN (a) and SVM (b) classi ie on LBP ea u es.
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
25
30
35
40
45
50
55
60
65
70
75
Accu acy
1-NN classi ie on GIST ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(a) 1-NN.
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
30
40
50
60
70
80
90
100
Accu acy
SVM classi ie on GIST ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(b) SVM.
Fig. 4: 1-NN (a) and SVM (b) classi ie on GIST ea u es.
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
45
50
55
60
65
70
75
80
Accu acy
1-NN classi ie on HOG ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(a) 1-NN.
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
65
70
75
80
85
90
95
100
Accu acy
SVM classi ie on HOG ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(b) SVM.
Fig. 5: 1-NN (a) and SVM (b) classi ie on HOG ea u es.
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 202

DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
25
30
35
40
45
50
55
60
65
70
75
Accu acy
1-NN classi ie on LBP + GIST ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(a) 1-NN.
0 10 20 30 40 50 60 70 80 90 100
Numbe o selec ed ea u es
30
40
50
60
70
80
90
100
Accu acy
SVM classi ie on LBP + GIST ea u es
CFS
FISHER
ILFS
LASSO
MRMR
RELIEFF
(b) SVM.
Fig. 6: 1-NN (a) and SVM (b) classi ie on LBP + GIST ea u es.
Tab. 2: LBP ea u es - classi ica ion pe o mance based on di e en ea u e selec ion me hods wi h 1-NN and SVM classi ie .
ACC: accu acy, Dim: dimension, id %: pe cen age o selec ed ea u es, ⩾id %: pe cen age o selec ed ea u es wi h
accu acy equal o all ea u es a e used.
LBP Dim
1-NN SVM
ACC Max ACC ACC Max ACC
100 % ⩾id % Dim Max id % Dim 100 % ⩾id % Dim Max id % Dim
Fishe 768 53.0 80 614 53.6 96 737 77.0 84 645 77.4 87 668
mRMR 768 53.0 11 84 59.0 28 215 77.0 22 169 81.8 37 284
Relie F 768 53.0 74 568 54.2 85 653 77.0 97 745 77.0 97 745
Il s 768 53.0 12 92 58.4 43 330 77.0 19 146 81.6 40 307
C s 768 53.0 90 691 52.3 96 737 77.0 96 737 77.1 96 737
Lasso 768 53.0 94 722 53.1 94 722 77.0 100 768 77.0 100 768
Tab. 3: LBP + GIST ea u es - classi ica ion pe o mance based on di e en ea u e selec ion me hods wi h 1-NN and SVM
classi ie . ACC: accu acy, Dim: dimension, id %: pe cen age o selec ed ea u es, ⩾id %: pe cen age o selec ed
ea u es wi h accu acy equal o all ea u es a e used.
LBP
Dim
1-NN SVM
+ ACC Max ACC ACC Max ACC
GIST 100 % ⩾id % Dim Max id % Dim 100 % ⩾id % Dim Max id % Dim
Fishe 1280 70.5 88 1126 70.7 88 1126 91.7 100 1280 91.7 100 1280
mRMR 1280 70.5 31 397 72.7 52 666 91.7 40 512 92.4 69 883
Relie F 1280 70.5 49 627 73.8 68 870 91.7 94 1203 91.9 96 1229
Il s 1280 70.5 27 346 72.4 72 922 91.7 41 525 94.2 58 742
C s 1280 70.5 59 755 70.9 94 1203 91.7 98 1254 91.7 98 1254
Lasso 1280 70.5 10 128 70.9 10 128 91.7 98 1254 91.7 98 1254
Tab. 4: LGIST ea u es - classi ica ion pe o mance based on di e en ea u e selec ion me hod wi h 1-NN and SVM classi ie .
ACC: accu acy, Dim: dimension, id %: pe cen age o selec ed ea u es, ⩾id %: pe cen age o selec ed ea u es wi h
accu acy equal o all ea u es a e used.
Dim GIST
1-NN SVM
ACC Max ACC ACC Max ACC
100 % ⩾id % Dim Max id % Dim 100 % ⩾id % Dim Max id % Dim
Fishe 512 69.4 42 215 70.2 47 241 88.3 98 502 88.3 98 502
mRMR 512 69.4 39 200 71.4 53 271 88.3 48 246 90.8 66 338
Relie F 512 69.4 21 108 73.4 70 358 88.3 36 184 90.2 46 236
Il s 512 69.4 49 251 70.0 79 404 88.3 99 507 88.4 99 507
C s 512 69.4 38 195 71.2 75 384 88.3 49 251 90.2 82 420
Lasso 512 69.4 40 205 69.7 99 507 88.3 58 297 90.6 78 399
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 203
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
Tab. 5: HOG ea u es - classi ica ion pe o mance based on di e en ea u e selec ion me hod wi h 1-NN and SVM classi ie .
ACC: accu acy, Dim: dimension, id %: pe cen age o selec ed ea u es, ⩾id %: pe cen age o selec ed ea u es wi h
accu acy equal o all ea u es a e used.
HOG Dim
1-NN SVM
ACC Max ACC ACC Max ACC
100 % ⩾id % Dim Max id % Dim 100 % ⩾id % Dim Max id % Dim
Fishe 21384 71.5 20 4277 73.2 27 5774 94.8 85 18176 94.8 99 21170
mRMR 21384 71.5 8 1711 73.9 14 2994 94.8 100 21384 94.8 100 21384
Relie F 21384 71.5 2 428 74.4 3642 94.8 100 21384 94.8 100 21384
Il s 21384 71.5 100 21384 71.5 100 21384 94.8 100 21384 94.8 100 21384
C s 21384 71.5 8 1711 72.9 21 4491 94.8 51 10906 95.1 74 15824
Lasso 21384 71.5 9 1925 75.5 19 4063 94.8 100 21384 94.8 100 21384
Tab. 6: The classi ica ion esul s ob ained by single and ensemble ea u e selec ion.
Classi ie
Da ase Single FS Mul i FS
Desc ip ion Dim ACC ACC max Acc Dim Pai Dim Ranke
ull wi hou FS o FSs (%) ull
(%) (%)
1-NN
LBP 768 53.0 59.0 60.0 432 Il s, 983 mRMR
Relie F
GIST 512 69.4 73.0 74.6 261 mRMR 655 Relie F
C s
HOG 21384 71.5 75.5 79.3 3416 mRMR 3635 Relie F
Relie F
LBP + GIST 1280 70.5 73.8 77.1 698 mRMR 1587 Relie F
Il s
SVM
LBP 768 77.0 81.8 82.4 544 Il s 591 mRMR
mRMR
GIST 512 88.3 90.8 91.4 1076
Il s
1346 mRMRmRMR
Fishe
LBP + GIST 1280 91.7 94.2 94.0 1246 mRMR 2112 Il s
Relie F
So, we combine wo bes subse o ea u es de e -
mined by Relie F and ILFS wi h a ea u e space equal
o 983. Nex , his ec o is applied again by mRMR
me hod and 1-NN classi ie o emo e i ele an
ea u es. Table 6 p esen s he compa ison o a single
and ensemble ea u e selec ion amewo k. We obse e
ha he ensemble me hod ou pe o ms single ea u e
selec ion me hod o all kinds o ea u es wi h 1-NN
classi ie . Fo example, we inc ease 1 % o accu acy
compa ed o a single ea u e selec ion me hod and in-
c ease 7 % compa ed wi h he classi ica ion when no
selec ion me hod is applied. Simila expe imen al e-
sul s a e ob ained by using SVM classi ie on single
iew desc ip o . In e ms o dimension, we inc ease
he ea u e space by combining and selec ing use ul in-
o ma ion o di e en single ea u e selec ion me hods.
Compa ed wi h he aims based on accu acy o ime
compu ing, an app op ia e app oach o such demand
has o be chosen.
5. Conclusion
In his pape , we in oduced a new ensemble ea u e
selec ion app oach by combining mul iple single ea-
u e selec ion me hods. A p e-selec ed subse o ea-
u es is i s de e mined by conside ing ea u e selec-
ion and associa ed classi ie . Mul iple subse s a e hen
combined o o m a inal ea u e space and hen ap-
plied ea u e selec ion me hod again o elimina e noisy
and edundan ea u es. The expe imen al esul s on
he VNRICE da ase o ice seed images classi ica ion
ha e shown he e iciency o he p oposed app oach.
The u u e o his wo k is o de e mine an app op i-
a e selec ion me hod based on each a ibu e and using
di e en s a egies o combine he inal ea u e ec o
esul ing om a single ea u e selec ion me hod.
Acknowledgmen
This wo k was suppo ed by Suan Sunandha Rajabha
Uni e si y, Thailand
Re e ences
[1] BENABDESLEM, K. and M. HINDAWI. Con-
s ained Laplacian Sco e o Semi-supe ised Fea-
u e Selec ion. In: Join Eu opean Con e ence on
Machine Lea ning and Knowledge Disco e y in
Da abases. Be lin: Sp inge , 2011, pp. 204–218.
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 204
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
ISBN 978-3-642-23780-5. DOI: 10.1007/978-3-642-
23780-5_23.
[2] BISHOP, C. M. Neu al Ne wo ks o Pa e n
Recogni ion. 1s ed. Ox o d: Ox o d uni e si y
p ess, 1996. ISBN 978-0-198-53864-6.
[3] BOLON-CANEDO, V. and A. ALONSO-
BETANZOS. Ensembles o ea u e selec ion:
A e iew and u u e ends. In o ma ion Fusion.
2019, ol. 52, iss. 1, pp. 1–12. ISSN 1566-2535.
DOI: 10.1016/j.in us.2018.11.008.
[4] CAI, J., J. LUO, S. WANG and S. YANG.
Fea u e selec ion in machine lea ning:
A new pe spec i e. Neu ocompu ing. 2018,
ol. 300, iss. 1, pp. 70–79. ISSN 0925-2312.
DOI: 10.1016/j.neucom.2017.11.077.
[5] CHIEW, K. L., C. L. TAN, K. WONG,
K. S. C. YONG and W. K. TIONG. A new
hyb id ensemble ea u e selec ion ame-
wo k o machine lea ning-based phishing
de ec ion sys em. In o ma ion Sciences. 2019,
ol. 484, iss. 1, pp. 153–166. ISSN 0020-0255.
DOI: 10.1016/j.ins.2019.01.064.
[6] DALAL, N. and B. TRIGGS. His og ams o o i-
en ed g adien s o human de ec ion. In: 2005
IEEE compu e socie y con e ence on compu e
ision and pa e n ecogni ion (CVPR’05). San
Diego: IEEE, 2005, pp. 886–893. IEEE. ISBN 0-
7695-2372-2. DOI: 10.1109/CVPR.2005.177.
[7] DENIZ, O., G. BUENO, J. SALIDO and
F. DE LA TORRE. Face ecogni ion us-
ing His og ams o O ien ed G adien s. Pa -
e n Recogni ion Le e s. 2011, ol. 32,
iss. 12, pp. 1598–1603. ISSN 1598-1603.
DOI: 10.1016/j.pa ec.2011.01.004.
[8] DROTAR, P., M. GAZDA and L. VOKOROKOS.
Ensemble ea u e selec ion using elec ion me h-
ods and anke clus e ing. In o ma ion Sciences.
2019, ol. 480, iss. 1, pp. 365–380. ISSN 0020-0255.
DOI: 10.1016/j.ins.2018.12.033.
[9] DUONG, H.-T. and V. T. HOANG. Dimensional-
i y Reduc ion Based on Fea u e Selec ion o Rice
Va ie ies Recogni ion. In: 2019 4 h In e na ional
Con e ence on In o ma ion Technology (InCIT).
Bangkok: IEEE, 2019, pp. 199–202. ISBN 978-1-
7281-1019-6. DOI: 10.1109/INCIT.2019.8912121.
[10] GOMES, J. F. S. and F. R. LETA. Applica-
ions o compu e ision echniques in he ag i-
cul u e and ood indus y: a e iew. Eu-
opean Food Resea ch and Technology. 2012,
ol. 235, iss. 6, pp. 989–1000. ISSN 1438-2385.
DOI: 10.1007/s00217-012-1844-2.
[11] GUYON, I. and A. ELISSEEFF. An In o-
duc ion o Va iable and Fea u e Selec ion.
Jou nal o Machine Lea ning Resea ch. 2003,
ol. 3, iss. 7, pp. 1157–1182. ISSN 1533-7928.
DOI: 10.5555/944919.944968.
[12] D. P. VAN HOAI, T. SURINWARANGKOON,
V. T. HOANG, H.-T. DUONG and
K. MEETHONGJAN. A Compa a i e S udy
o Rice Va ie y Classi ica ion based on Deep
Lea ning and Hand-c a ed Fea u es. ECTI
T ansac ions on Compu e and In o ma ion
Technology (ECTI-CIT). 2020, ol. 14, iss. 1,
pp. 1–10. ISSN 2286-9131. DOI: 10.37936/ec i-
ci .2020141.204170.
[13] HUMEAU-HEURTIER, A. Tex u e Fea u e Ex-
ac ion Me hods: A Su ey. IEEE Access. 2019,
ol. 7, iss. 1, pp. 8975–9000. ISSN 2169-3536.
DOI: 10.1109/ACCESS.2018.2890743.
[14] KIRA, K. and L. A. RENDELL. A P ac i-
cal App oach o Fea u e Selec ion. In: Ma-
chine Lea ning P oceedings 1992. Abe deen: Else-
ie , 1992, pp. 249–256. ISBN 978-1-558-60247-2.
DOI: 10.1016/B978-1-55860-247-2.50037-1.
[15] KONONENKO, I. Es ima ing a ibu es: Anal-
ysis and ex ensions o RELIEF. In: Eu o-
pean Con e ence on Machine Lea ning. Sp inge ,
1994, pp. 171–182. ISBN 978-3-540-48365-6.
DOI: 10.1007/3-540-57868-4_57.
[16] KONONENKO, I., E. SIMEC and M. ROBNIK-
SIKONJA. O e coming he Myopia o Induc-
i e Lea ning Algo i hms wi h RELIEFF. Ap-
plied In elligence. 1997, ol. 7, iss. 1, pp. 39–55.
ISSN 1573-7497. DOI: 10.1023/A:1008280620621.
[17] MEBATSION, H. K., J. PALIWAL and
D. S. JAYAS. Au oma ic classi ica ion o non-
ouching ce eal g ains in digi al images using
limi ed mo phological and colo ea u es. Com-
pu e s and Elec onics in Ag icul u e. 2013,
ol. 90, iss. 1, pp. 99–105. ISSN 0168-1699.
DOI: 10.1016/j.compag.2012.09.007.
[18] MIFTAHUSHUDUR, T., C. B. A. WAEL and
T. PRALUDI. In ini e La en Fea u e Selec-
ion Technique o Hype spec al Image Classi-
ica ion. Ju nal Elek onika dan Telekomunikasi.
2019, ol. 19, iss. 1, pp. 32–37. ISSN 2527-9955.
DOI: 10.14203/je . 19.32-37.
[19] MIRZAEI, A., V. POURAHMADI, M. SOLTANI
and H. SHEIKHZADEH. Deep ea u e selec ion
using a eache -s uden ne wo k. Neu ocompu ing.
2020, ol. 383, iss. 1, pp. 396–408. ISSN 0925-2312.
DOI: 10.1016/j.neucom.2019.12.017.
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 205
DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 18 |NUMBER: 3 |2020 |SEPTEMBER
[20] NHAT, H. T. M. and V. T. HOANG. Fea-
u e usion by using LBP, HOG, GIST desc ip-
o s and Canonical Co ela ion Analysis o ace
ecogni ion. In: 2019 26 h In e na ional Con-
e ence on Telecommunica ions (ICT). Hanoi:
IEEE, 2019, pp. 371–375. ISBN 978-1-7281-0273-
3. DOI: 10.1109/ICT.2019.8798816.
[21] OJALA, T., M. PIETIKAINEN and T. MAEN-
PAA. A Gene alized Local Bina y Pa e n Op-
e a o o Mul i esolu ion G ay Scale and Ro a-
ion In a ian Tex u e Classi ica ion. In: In e na-
ional Con e ence on Ad ances in Pa e n Recog-
ni ion. Rio de Janei o: Sp inge , 2001, pp. 399–
408. ISBN 978-3-540-41767-5. DOI: 10.1007/3-
540-44732-6_41.
[22] OLIVA, A. and A. TORRALBA. Modeling
he Shape o he Scene: A Holis ic Rep-
esen a ion o he Spa ial En elope. In e -
na ional Jou nal o Compu e Vision. 2001,
ol. 42, iss. 3, pp. 145–175. ISSN 0920-5691.
DOI: 10.1023/A:1011139631724.
[23] SEIJO-PARDO, B., I. PORTO-DIAZ,
V. BOLON-CANEDO and A. ALONSO-
BETANZOS. Ensemble ea u e selec ion:
Homogeneous and he e ogeneous app oaches.
Knowledge-Based Sys ems. 2017, ol. 118,
iss. 1, pp. 124–139. ISSN 0950-7051.
DOI: 10.1016/j.knosys.2016.11.017.
[24] VAN, T. N. and V. T. HOANG. Kinship Ve i ica-
ion based on Local Bina y Pa e n ea u es cod-
ing in di e en colo space. In: 2019 26 h In e na-
ional Con e ence on Telecommunica ions (ICT).
Hanoi: IEEE, 2019, pp. 376–380. ISBN 978-1-
7281-0273-3. DOI: 10.1109/ICT.2019.8798781.
[25] YAMADA, M., W. JITKRITTUM, L. SI-
GAL, E. P. XING and M. SUGIYAMA. High-
Dimensional Fea u e Selec ion by Fea u e-Wise
Ke nelized Lasso. Neu al Compu a ion. 2014,
ol. 26, iss. 1, pp. 185–207. ISSN 1530-888X.
DOI: 10.1162/NECO_a_00537.
[26] ZHANG, R., F. NIE, X. LI and X. WEI.
Fea u e selec ion wi h mul i- iew da a:
A su ey. In o ma ion Fusion. 2019, ol. 50,
iss. 1, pp. 158–167. ISSN 1566-2535.
DOI: 10.1016/j.in us.2018.11.019.
[27] ZHAO, Z., R. ANAND and M. WANG. Max-
imum Rele ance and Minimum Redundancy
Fea u e Selec ion Me hods o a Ma ke ing
Machine Lea ning Pla o m. In: 2019 IEEE
In e na ional Con e ence on Da a Science and Ad-
anced Analy ics (DSAA). Washing on: IEEE,
2019, pp. 442–452. ISBN 978-1-7281-4493-1.
DOI: 10.1109/DSAA.2019.00059.
Abou Au ho s
Dzi Lam T an TUAN is a g adua e s uden
o Ho Chi Minh Ci y Open Uni e si y, Vie nam.
He ob ained his M.Sc. deg ee in Compu e Science in
2019. His esea ch in e es s include ea u e selec ion
and image analysis.
Thongchai SURINWARANGKOON was bo n
in Su a hani, Thailand in 1972. He ob ained a B.Sc.
deg ee in ma hema ics om Chiang Mai Uni e si y,
Thailand in 1995. He ecei ed M.Sc. deg ee in
managemen o in o ma ion echnology om Walailak
Uni e si y, Thailand in 2005 and Ph.D. in in o -
ma ion echnology om King Mongku ’s Uni e si y
o Technology No h Bangkok, Thailand in 2014.
He is now a Lec u e in he Depa men o Business
Compu e , Suan Sunandha Rajabha Uni e si y,
Bangkok, Thailand. His esea ch in e es s include
digi al image p ocessing, machine lea ning, a i icial
in elligence, business in elligence, and mobile applica-
ion de elopmen o business.
Ki ikhun MEETHONGJAN is a ull lec u e in
he Compu e Science P og am, Head o Apply Science
Depa men , Facul y o Science and Technology, Suan
Sunandha Rajabha Uni e si y (SSRU), Thailand.
He ecei ed his B.Sc. Deg ee in Compu e Science
om SSRU in 1990, a M.Sc. deg ee in compu e
science om King Mongku ’s Uni e si y o Technology
Thonbu i (KMUTT) in 2000 and a Ph.D. deg ee in
Compu e G aphic om he Uni e si y o Technology,
Malaysia, in 2013. His esea ch in e es is Compu e
G aphic, Image P ocessing, A i icial In elligen ,
Biome ics, Pa e n Recogni ion, So Compu ing
echniques and, applica ion. Cu en ly, he is an ad i-
so o Mas e and Ph.D. s uden in Fo ensic Science
o SSRU.
Vinh T uong HOANG ecei ed his M.Sc. de-
g ee om he Uni e si y o Mon pellie in 2009 and his
Ph.D. deg ee in compu e science om he Uni e si y
o he Li o al Opal Coas , F ance. He is cu en ly
an assis an p o esso and Head o Image P ocessing
and Compu e G aphics Depa men a he Ho Chi
Minh Ci y Open Uni e si y, Vie nam. His esea ch
in e es s include image analysis and ea u e selec ion.
c
2020 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 206