Inter
national
J
our
nal
of
Inf
ormatics
and
Communication
T
echnology
(IJ-ICT)
V
ol.
15,
No.
4,
December
2026,
pp.
1861
∼
1870
ISSN:
2252-8776,
DOI:
10.11591/ijict.v15i4.pp1861-1870
❒
1861
A
muzzle-based
self-enhancing
co
w
r
ecognition
method
using
Y
OLO
and
LLM
Rifat
Sark
er
Aoy
on
1
,
Ismail
Hossain
2
,
Jia
Uddin
3
1
Department
of
Computer
Science
and
Engineering,
Brac
Uni
v
ersity
,
Dhaka,
Bangladesh
2
Department
of
Computer
Science
and
Engineering,
Geor
ge
Mason
Uni
v
ersity
,
F
airf
ax,
USA
3
Department
of
AI
and
Big
Data,
Endicott
Colle
ge,
W
oosong
Uni
v
ersity
,
Daejeon,
South
K
orea
Article
Inf
o
Article
history:
Recei
v
ed
Dec
30,
2024
Re
vised
Jun
10,
2026
Accepted
Jul
5,
2026
K
eyw
ords:
Co
w
muzzle
Image
processing
LLM
Self-enhancing
Y
OLO
ABSTRA
CT
In
this
research,
we
ha
v
e
presented
an
approach
to
muzzle-based
co
w
identica-
tion.
T
o
recognise
muzzles,
we
ha
v
e
used
you
only
look
once
(Y
OLO)
and
the
structural
similarity
inde
x
(SSI).
In
the
proposed
method,
a
lar
ge
amount
of
data
is
not
required
for
the
image
processing
task.
Because
here
we
included
a
self-
enhancing
feature,
which
helps
the
model
to
predict
accurately
and
also
helps
to
impro
v
e
performance
by
increasing
the
amount
of
data.
This
certain
program
helps
the
model
to
impro
v
e
its
capabi
lity
for
deli
v
ering
f
aster
responses.
lar
ge
language
model
(LLM)
is
responsible
for
adding
the
self-enhancing
feature.
On
the
other
ha
nd,
to
nd
out
the
m
ost
suitable
Y
OLO
model
for
our
research,
we
ha
v
e
trained
data
with
three
dif
ferent
v
ersions
of
Y
OLO:
Y
OLO
v7,
v8,
and
v9.
W
e
e
v
aluated
the
performances
of
these
three
v
ersions
in
terms
of
Preci-
sion,
Recall,
mAP50,
and
mAP50-95
and
found
Y
OLO
v9
to
be
the
best
among
them.
It
e
xcels
with
a
precisi
on
of
0.391,
a
recall
of
0.916,
a
mAP50
of
0.559,
and
a
mAP50-95
of
0.307.
Here,
data
ha
v
e
been
collect
ed
manually
and
were
annotated
by
Roboo
w
.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
Jia
Uddin
Department
of
AI
and
Big
Data,
Endicott
Colle
ge,
W
oosong
Uni
v
ersity
Daejeon,
South
K
orea
Email:
jia.uddin@wsu.ac.kr
1.
INTR
ODUCTION
The
co
w
is
one
of
the
most
essential
domestic
animals
in
the
entire
w
orld.
People
dom
esticate
co
ws
for
v
arious
reasons,
lik
e
the
production
of
dairy
,
meat,
leather
accessories,
and
so
on
[1].
Moreo
v
er
,
co
ws
can
play
crucial
roles
in
the
economy
of
a
re
gion
[2].
So,
people
mak
e
co
w
f
arms
so
that
the
y
can
tak
e
care
of
co
ws
appropriately
.
T
o
or
g
anize
properly
,
sometimes
co
ws
need
to
tak
e
treatment.
In
this
instance,
co
w
recognition
is
highl
y
needed.
Besides,
there
are
se
v
eral
re
gions
of
the
w
orld
where
the
f
armer
can
get
insurance
for
the
co
w
[3].
In
this
case,
co
w
recognition
is
also
required.
Lik
e
these
tw
o
sectors,
there
are
also
man
y
more
sectors
a
v
ailable
where
co
w
recognition
systems
can
play
v
ery
important
roles
[4].
There
are
man
y
AI-based
co
w
recognition
research
projects
that
ha
v
e
been
conducted
till
no
w
.
In
AI-based
solutions,
data
is
pretty
essential.
On
the
other
hand,
it
is
time
-consuming
and
pretty
complicated
to
collect
a
suf
cient
amount
of
data.
Additionally
,
the
AI-based
architectures
are
not
suitable
for
non-computer
scientists.
Moreo
v
er
,
when
the
model
encounters
data
which
is
not
f
amiliar
to
the
model,
then
it
will
be
dif
cul
t
for
the
model
to
pro
vide
appropriate
output.
That
is
wh
y
we
ha
v
e
decided
to
mak
e
a
no
v
el
frame
w
ork
of
co
w
recognition
that
does
not
need
a
lar
ge
amount
of
data
and
is
easy
for
an
yone
to
use,
e
v
en
without
a
background
J
ournal
homepage:
http://ijict.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
1862
❒
ISSN:
2252-8776
in
computer
science.
W
e
used
you
only
look
once
(Y
OLO)
and
structura
l
similarity
inde
x
(SSI)
to
recognize
co
ws
[5].
Y
OLO
is
pretty
signicant
for
image
processing
tasks
lik
e
object
detection,
image
se
gmentation,
and
so
on.
Besides,
for
real-time
predictions,
the
performance
of
Y
OLO
is
quite
phenomenal.
As
a
consequence,
we
decided
to
use
Y
OLO
for
detecting
muzzles.
Afterw
ards,
we
completed
the
task
of
muzzle
comparison
via
SSI.
The
lar
ge
language
model
(LLM)
is
trained
on
v
ast
datasets
to
understand
and
respond
to
human
language
using
transformer
architecture
[6].
Y
OLO
has
launched
se
v
eral
v
ersions
till
no
w
.
Y
OLO
v7
is
one
of
the
amazing
v
ersions
of
Y
OLO
for
image
processing
and
computer
vision
tasks
[7].
Y
OLO
v7
is
comparati
v
ely
better
than
its
predecessors
in
terms
of
accurac
y
of
real-time
object
detection
and
reduced
40%
parameter
than
the
earlier
edition
of
the
Y
OLO.
W
e
can
di
vide
the
architecture
of
Y
OLO
v7
into
mainly
three
se
gments,
which
are
the
backbone,
neck,
and
head.
Here,
the
backbone
is
responsible
for
fetching
essential
features
from
the
images
and
sending
these
to
the
Neck.
The
Neck
combines
all
the
features
of
the
backbone
and
mak
es
a
p
yramid.
Finally
,
Head
w
orks
for
output
prediction.
Y
OLO
v8
is
one
of
the
v
ersions
of
Y
OLO
that
can
outperform
most
other
image
processing
algorithms
[8].
T
o
predict
the
bounding
box
or
an
y
class
from
the
Y
OLO
v8
model,
it
is
required
to
follo
w
se
v
eral
steps.
In
this
case,
the
rst
one
is
di
viding
the
image
into
a
grid.
The
ne
xt
task
is
to
e
xtract
features
using
con
v
olutional
neural
netw
ork
(CNN)
[9].
Then,
the
Y
OLO
v8
model
can
predict
the
bounding
box
or
class.
And
In
Y
OLO
v9,
programmable
gradient
information
(PGI)
and
the
generalized
ef
cient
layer
aggre
g
ation
netw
ork
(GELAN)
enhanced
its
performance
[10].
Lik
e
Y
OLO
v7,
the
architecture
of
Y
OLO
v9
can
be
di
vided
into
three
parts:
the
backbone,
neck,
and
head.
Furthermore,
the
acti
vities
of
these
three
parts
are
also
similar
.
But
Y
OLO
v9
has
an
adv
anced
layer
,
which
enables
Y
OLO
v9
to
perform
better
.
In
this
research,
we
trained
model
with
three
dif
ferent
v
ersions
of
Y
OLO
and
pro
vided
their
performance.
It
will
assist
us
in
getting
an
idea
of
the
appropriate
v
ersion
of
Y
OLO
for
our
proposed
research.
Along
with
these,
we
ha
v
e
added
self-enhancing
features
that
enable
our
proposed
system
to
con-
sistently
de
v
elop
performance
by
interacting
with
the
users.
T
o
de
v
elop
this
no
v
el
feature,
we
used
LLM
technology
[11].
Our
moti
v
e
in
this
research
is
to
launch
a
platform
that
will
be
capable
of
recognizing
co
ws
and
impro
ving
the
performance
of
the
platform
consistently
by
interacting
with
its
users.
This
co
w
recognition
system
can
be
di
vided
into
tw
o
major
se
gments,
one
of
which
is
detecting
muzzles
from
the
co
w
image,
and
the
other
is
comparing
muzzles.
W
e
ha
v
e
used
Y
OLO
and
SSI
for
the
tasks
of
muzzle
detection
and
muzzle
comparison,
respecti
v
ely
.
B
ecause,
Y
OLO
models
are
v
ery
popular
for
deplo
yment
because
the
y
are
f
ast,
easy
to
use,
w
ork
well
on
lo
w-cost
de
vices,
and
ha
v
e
good
support
from
the
open-source
community
.
So,
Y
OLO
is
quite
essential
for
non-computer
scientists,
who
w
ant
to
deplo
y
object
detection
models
in
real-w
orld
applica-
tions.
On
the
other
hand,
LLM
technology
is
used
here
to
impro
v
e
performance
by
pro
viding
a
f
aster
response.
Besides,
LLM
will
help
increase
the
amount
of
data
consistently
by
interacting
with
the
users.
Also,
we
ha
v
e
tested
the
performance
of
three
v
ersions
of
Y
OLO,
named
Y
OLO
v7,
v8,
and
v9
to
nd
the
most
suitable
v
ersions
for
our
research.
2.
RELA
TED
W
ORKS
Zin
et
al.
[12]
ha
v
e
presented
an
approach
to
identify
the
co
w
based
on
the
ear
tag.
Here,
to
fulll
this
task
appropriately
,
we
used
Y
OLO
and
CNN.
The
y
ha
v
e
collected
data
from
t
he
videos
of
co
ws
on
a
dairy
f
arm
in
Japan.
Lu
et
al.
[13]
ha
v
e
disclosed
a
w
ay
of
identifying
co
ws
by
analyzing
their
iris.
The
y
analyzed
the
iris
in
three
steps.
Here,
the
last
part
is
co
w
recognition,
and
transform
(2D-CWT)
is
used
by
them
to
do
this
recognizing
task.
Gu
et
al.
[14]
ha
v
e
presented
a
w
ay
to
recognize
the
beha
vior
of
the
co
ws
in
terms
of
their
acti
vities.
T
o
accomplish
this
task,
the
y
collected
400
data
of
the
co
ws.
Y
ao
et
al.
[15]
ha
v
e
shared
an
idea
of
recognizing
the
co
w
f
ace
using
feature
e
xtraction.
In
this
research,
the
authors
ha
v
e
used
50,000
data
and
got
accurac
y
of
98.3%
and
94.1%
for
the
detection
and
co
w
f
ace
recognition,
respecti
v
ely
.
Li
et
al.
[16]
proposed
an
approach
to
recognise
co
ws
using
CNN.
Here,
the
authors
ha
v
e
de
v
eloped
the
CNN
by
including
Ale
xnet.This
impro
v
ed
model
is
able
to
secure
the
accurac
y
of
97.95%.
Zin
et
al.
[17]
ha
v
e
disclosed
a
unique
w
ay
to
recognize
co
ws
in
terms
of
the
black
and
white
pattern
of
their
bodies.
Initially
,
the
y
used
inter
-frame
dif
ferencing
and
a
horizontal
histogram-based
approach
to
fetch
the
body
from
the
co
w
image
and
cropped
this
se
gment
afterw
ards.
Ev
entually
,
this
cer
-
tain
re
gion
w
as
utilized
for
training
via
deep
learning
algorithm.
K
umar
et
al.
[18]
ha
v
e
sho
wn
a
w
ay
to
identify
the
co
ws
based
on
the
muzzle.
The
y
di
vided
the
entire
research
into
mainly
tw
o
se
gments,
where
the
rst
is
detecting
muzzles
and
the
last
is
comparing
muzzles
with
the
database.
The
y
used
machine
learn-
ing
algorithm
for
detecting
muzzles
from
the
image
of
the
co
w
f
ace.
On
the
other
hand,
one-shot-similarity
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
4,
December
2026:
1861–1870
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
1863
(OSS)
matching
and
distance
metric
learning
based
techni
ques
were
used
for
comparing
the
muzzle
structure.
Bello
et
al.
[19]
ha
v
e
in
v
ented
an
incredible
w
ay
to
identify
co
ws
using
their
body
patterns.
The
y
used
CNN
to
train
the
model
for
co
w
recognition,
while
the
amount
of
data
w
as
1,000
images
of
the
co
ws.
Zin
et
al.
[20]
ha
v
e
presented
a
method
of
co
w
tracking
using
the
ear
tag
of
the
co
w
.
At
the
be
ginning,
the
y
concentrated
on
detecting
heads
and
did
this
via
Y
OLO.
Afterw
ards,
the
y
w
ork
ed
on
ear
recognition
and
ha
v
e
d
one
so
in
three
se
gments.
T
o
fulll
these
three
parts,
the
y
used
image
processing
and
CNN
technologies.
Bello
et
al.
[21]
ha
v
e
sho
wn
a
method
to
gure
out
the
co
w
nose
pattern
.
T
o
fulll
this
job,
the
y
used
Stack
ed
denoising
auto-encoder
and
deep
belief
netw
ork.
T
o
train
the
model,
the
y
used
data
from
4,000
images.
K
umar
et
al.
[22]
proposed
an
approach
to
recognise
the
f
aces
of
cattle
using
speeded
up
rob
ust
feature
and
local
binary
patterns.
The
y
used
1,200
images
to
train
the
model
and
got
accurac
y
of
92.5%.
W
ang
et
al.
[23]
ha
v
e
sho
wn
a
method
for
identifying
estrus
co
ws
using
Y
OLOv8n,
where
the
y
used
1716
data.
Besides,
the
y
ha
v
e
also
highlighted
the
ef
fecti
v
eness
of
their
proposed
model.
Li
et
al.
[24]
used
a
phenomenal
dataset
to
identify
co
w
beha
vior
.
The
y
ha
v
e
collected
data
from
107
co
ws
and
di
vided
their
beha
vior
into
v
e
classes.
W
eng
et
al.
[25]
introduced
a
unique
CNN
to
recognize
co
ws.
All
the
data
were
collected
from
tw
o
co
ws,
and
images
were
collected
from
dif
ferent
angles
to
train
properly
.
K
umar
et
al.
[26]
ha
v
e
presented
a
frame
w
ork
via
Y
OLO
to
recognize
the
co
w
.
Here,
the
propesed
framwe
w
ok
w
orks
to
fetch
only
important
feature
of
co
w
f
ace
and
muzzle
for
ensuring
both
resource
ef
cienc
y
and
acc
u
r
ate
cattle
identication.
Bakhshayeshi
et
al.
[27]
ha
v
e
presented
a
method
to
recognise
co
ws
le
v
eraging
Y
OLOv5
and
Siamese
neural
netw
orks.
W
ith
only
2,500
co
w
images,
their
proposed
system
secures
signicant
accurac
y
.
Additionally
,
this
frame
w
ork
has
the
scope
to
inte
grate
v
arious
sectors
lik
e
f
arms,
car
go
areas,
sale
yards
and
so
on.
3.
METHOD
T
o
compl
ete
our
research,
we
ha
v
e
follo
wed
the
sequence
in
Figure
1.
Initially
,
we
concentrated
on
collecting
data
and
did
the
remaining
tasks
afterw
ards.
Figure
1.
Process
sequence
of
our
research
3.1.
Pr
eparing
data
W
e
initially
collected
the
images
of
the
co
w
muzzle
from
online
sources,
some
of
t
hem
manually
in
real
life.
W
e
ha
v
e
collected
a
total
of
189
images
of
co
w
muzzles
[28].
Afterw
ards,
we
used
roboo
w
to
create
a
bounding
box
in
the
re
gion
of
the
co
w
muzzle.
Here,
142
data
were
allocated
for
training,
whereas
31
and
16
images
were
used
for
v
alidation
and
testing
purposes.
Each
image
w
as
resized
to
a
square
dimension
of
640x640
pix
els.
This
resizing
w
as
performed
by
stretching
the
image
to
t
the
specied
dimensions,
which
may
alter
the
original
aspect
ratio
of
the
image.
3.2.
T
raining
and
muzzle
detection
T
o
train
the
model
for
detecting
muzzles
from
the
co
w
images,
we
used
the
Y
OLO
model.
T
o
nd
out
the
best
Y
OLO
model,
we
trained
the
data
with
three
dif
ferent
v
ersions
of
Y
OLO.
These
are
Y
OLO
v7,
v8,
and
v9.
W
e
run
each
v
ersion
for
up
to
25
epochs.
In
T
able
1,
the
information
for
the
training
conguration
of
each
model
is
gi
v
en.
W
e
ha
v
e
also
pro
vided
the
three
model
summaries
in
T
able
1.
Aft
er
training
the
models,
the
models
were
sa
v
ed
in
the
PyT
orch
format
with
the.pt
e
xtension.
This
le,
which
contains
the
trained
model’
s
weights
and
architecture,
w
as
subsequently
utilized
for
detecting
muzzles.
W
e
get
the
prediction
of
the
coordinate
of
the
muzzle
from
the
trained
model.
Then,
the
muzzle
part
is
cropped
using
the
predicted
coordinates.
Afterw
ards,
this
cropped
re
gion
will
be
mo
v
ed
forw
ard
to
the
comparison
phase.
A
muzzle-based
self-enhancing
cow
r
eco
gnition
method
.
.
.
(Rifat
Sark
er
Aoyon)
Evaluation Warning : The document was created with Spire.PDF for Python.
1864
❒
ISSN:
2252-8776
T
able
1.
T
raining
conguration
and
model
summary
Model
name
Epochs
Image
size
Batch
size
Momentum
W
eight
decay
Layers
P
arameters
Gradients
GFLOPs
Y
OLOv7
25
640
×
640
16
0.937
0.0005
415
37,201,950
37,201,950
105.1
Y
OLOv8
25
800
×
800
16
0.937
0.0005
225
11,136,374
11,136,358
28.6
Y
OLOv9
25
640
×
640
16
0.937
0.0005
621
25,438,614
25,438,598
103.2
3.3.
Muzzle
comparision
At
rst,
the
tw
o
muzzles
will
be
con
v
erted
to
the
same
dimensions
and
con
v
erted
to
grayscale
after
-
w
ards.
F
ollo
wing
that,
the
absolute
dif
ference
between
the
grayscale
images
is
computed.
This
dif
ference
in
image
highlights
the
re
gions
where
pix
el
v
alues
v
ary
between
the
tw
o
images.
The
dif
ference
image
is
then
thresholded
to
produce
a
binary
image.
Pix
els
with
dif
ferences
abo
v
e
a
specied
threshold
are
mark
ed,
empha-
sizing
the
areas
of
signicant
change.
After
that,
the
SSI
is
calculated
[29].
The
SSI
measures
ho
w
similar
tw
o
grayscale
images
are
by
comparing
their
brightness,
contrast,
and
structure.
Grayscale
images
are
used
in
SSI
calculations
to
simplify
the
comparison
by
focusing
s
olely
on
brightness
and
structural
features
without
the
added
comple
xity
of
color
dif
ferences.
This
mak
es
it
easier
to
assess
and
quantify
similarities
and
dif
ferences
in
image
content.
If
the
SSI
v
alue
is
greater
than
0.66,
then
it
will
be
considered
that
tw
o
muzzles
are
the
same.
Otherwise,
these
are
dif
ferent.
3.3.1.
Structural
similarity
index
The
SSI
is
an
incredible
algorithm
to
measure
the
similarity
between
tw
o
images.
T
o
fetch
the
s
imilar
-
ity
via
SSI,
the
pix
el-wise
comparison
is
not
just
between
the
tw
o
images;
al
so,
the
o
v
erall
s
tructural,
contrast,
and
luminance
similarity
between
the
images
will
be
calculat
ed.
T
o
e
xtract
the
SSI,
we
rst
need
to
gure
out
the
dif
ference
between
tw
o
images
of
luminance,
contrast,
and
structure.
Luminance
calculation
is
the
process
of
calculating
the
a
v
erage
brightness
of
both
images.
On
the
other
hand,
contrast
calculation
means
measuring
the
v
ariations
in
brightness.
The
structure
dif
ference
is
the
consistenc
y
of
the
patterns
and
te
xtures
between
the
tw
o
images.
After
nding
these
three
v
alues
,
it
is
needed
to
combine
this
three
v
alues.Then
we
will
get
the
SSI
[30].
The
process
of
generating
SSI
is
visually
sho
wn
in
Figure
2.
Figure
2.
Process
of
generating
SSI
The
luminance
comparison
between
tw
o
images
x
and
y
is
gi
v
en
by
(1),
which
measures
ho
w
similar
the
a
v
erage
brightness
le
v
els
of
the
images
are.
Here,
µ
x
and
µ
y
represent
the
m
ean
luminance
of
images
x
and
y
,
respecti
v
ely
,
and
C
1
is
a
constant
to
a
v
oid
di
vision
by
zero.
The
constant
C
1
is
dened
in
(2).
l
(
x,
y
)
=
2
µ
x
µ
y
+
C
1
µ
2
x
+
µ
2
y
+
C
1
(1)
C
1
=
(
K
1
·
L
)
2
(2)
The
contrast
comparison
is
dened
by
(3),
which
e
v
aluates
the
similarity
in
contrast
between
images
x
and
y
.
Here,
σ
x
and
σ
y
denote
the
standard
de
viations
of
pix
el
v
alues
in
images
x
and
y
,
respecti
v
ely
,
and
C
2
is
a
constant
to
a
v
oid
di
vision
by
zero.
The
constant
C
2
is
dened
in
(4).
c
(
x,
y
)
=
2
σ
x
σ
y
+
C
2
σ
2
x
+
σ
2
y
+
C
2
(3)
C
2
=
(
K
2
·
L
)
2
(4)
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
4,
December
2026:
1861–1870
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
1865
In
(5)
describes
the
structure
comparison,
which
assesses
ho
w
similar
the
structural
patterns
in
images
x
and
y
are.
The
co
v
ariance
σ
xy
,
along
with
the
standard
de
viations
σ
x
and
σ
y
,
and
constant
C
3
,
which
is
often
dened
as
C
3
=
C
2
2
,
are
used
in
this
calculation.
s
(
x,
y
)
=
σ
xy
+
C
3
σ
x
σ
y
+
C
3
(5)
The
nal
SSI
is
gi
v
en
by
(6),
which
combines
the
luminance,
contrast,
and
structure
compari
sons
using
weights
α
,
β
,
and
γ
to
determine
the
o
v
erall
similarity
between
the
images.
S
S
I
(
x,
y
)
=
[
l
(
x,
y
)]
α
·
[
c
(
x,
y
)]
β
·
[
s
(
x,
y
)]
γ
(6)
Finally
,
the
co
v
ariance
calculation
i
s
pro
vided
in
(7).
This
equation
calculates
the
co
v
ariance
σ
xy
between
images
x
and
y
,
which
is
crucial
for
the
structure
comparison
component.
σ
xy
=
1
N
−
1
N
X
i
=1
(
x
i
−
µ
x
)(
y
i
−
µ
y
)
(7)
Precision
=
T
rue
Positi
v
es
T
rue
Positi
v
es
+
F
alse
Positi
v
es
(8)
Recall
=
T
rue
Positi
v
es
T
rue
Positi
v
es
+
F
alse
Ne
g
ati
v
es
(9)
mAP50
=
1
C
C
X
i
=1
AP
i
(
IoU
=
0
.
5
0)
(10)
mAP50-95
=
1
C
C
X
i
=1
1
T
T
X
t
=1
AP
i
(
IoU
t
)
(11)
3.4.
Self
enhancement
After
predicting
the
muzzle
from
the
Y
OLO
model,
it
will
rst
be
check
ed
to
see
whether
the
predic
ted
output
is
accurate
or
not.
T
o
do
so,
we
send
a
query
to
LLM
with
the
output
image.
If
the
result
is
positi
v
e,
the
process
will
be
forw
arded
to
the
ne
xt
step.
Otherwise,
the
accurate
coordinates
of
the
muzzle
bounding
box
will
be
predicted
via
LLM.
Afterw
ards,
the
image
and
predicted
coordinates
will
be
sa
v
ed
in
the
database
and
will
mo
v
e
on
to
the
ne
xt
phase.
Thus,
the
amount
of
data
will
be
consistently
increasing
by
interacting
with
the
users,
which
will
impro
v
e
the
performance
of
pro
viding
f
aster
responses.
Ho
we
v
er
,
this
platform
will
al
w
ays
be
capable
of
pro
viding
the
correct
response.
Here,
we
used
GPT
4o
Api
for
getting
output
[31].
T
o
get
an
accur
ate
prediction
from
GPT
4o
LLM,
we
initially
pro
vided
se
v
eral
co
w
images
with
their
corresponding
coordinates
of
the
muzzle
so
that
LLM
can
understand
the
pattern
of
recognition
of
the
re
gion
of
the
muzzle.
The
process
sequence
of
self-enhancement
feature
is
visually
gi
v
en
in
Figure
3.
Figure
3.
Process
sequence
of
Self
Enhancement
A
muzzle-based
self-enhancing
cow
r
eco
gnition
method
.
.
.
(Rifat
Sark
er
Aoyon)
Evaluation Warning : The document was created with Spire.PDF for Python.
1866
❒
ISSN:
2252-8776
3.5.
Ultimate
output
pr
ediction
The
o
v
ervie
w
of
ultimate
output
prediction
is
visually
presented
in
the
Figure
4.
Initially
,
the
user
needs
to
pro
vide
tw
o
inputs.
F
ollo
wing
that,
the
coordinates
of
the
muzzles
will
be
predicted
from
the
trained
model
for
tw
o
inputs
separately
.
Whether
the
prediction
is
accurate
or
not
will
be
check
ed
by
ou
r
proposed
“Self
Enhancement”
feature.
After
going
through
the
process
of
“Self
Enhancement”,
the
ultimate
output
of
muzzle
co-ordinate
will
be
predicted.
Then,
muzzles
will
be
cropped,
and
these
cropped
muzzles
will
be
compared
based
on
the
SSI.
If
the
SSI
v
alue
is
greater
than
0.66,
then
it
will
be
considered
that
tw
o
muzzles
are
the
same.
Otherwise,
these
are
dif
ferent.
Figure
4.
Ov
ervie
w
of
ultimate
output
prediction
4.
RESUL
T
AN
AL
YSIS
In
our
research,
we
trained
the
data
using
three
dif
f
erent
v
ersions
of
Y
OLO.
The
performance
of
Y
OLOv7,
Y
OLOv8,
and
Y
OLOv9
w
as
e
v
aluated
based
on
four
metrics:
precision,
recall,
mAP50,
and
mAP50-
95
[32]-[35].
The
results
of
these
three
v
ersions
in
terms
of
the
four
metrics
are
pro
vided
in
T
able
2.
Precision
is
the
accurac
y
of
positi
v
e
predictions,
while
recall
measures
the
ability
to
identify
all
rele
v
ant
instances
of
a
model.
The
formulas
for
precision
and
recall
are
pro
vided
in
(8)
and
(9),
respecti
v
ely
.
On
the
other
hand,
mAP50
is
the
mean
a
v
erage
precision
calculated
at
an
intersection
o
v
er
union
(IoU)
threshold
of
0.50,
while
mAP50-95
is
the
mean
a
v
erage
precision
a
v
eraged
o
v
er
multiple
IoU
thresholds
ranging
from
0.50
to
0.95.
The
formulas
for
mAP50
and
mAP50-95
are
gi
v
en
in
(10)
and
(11),
respecti
v
ely
.
From
T
able
2,
we
can
see
that
Y
OLOv9
outperforms
the
other
v
ersions
in
terms
of
eecall,
mAP50,
and
mAP50-95.
Ho
we
v
er
,
Y
OLOv7
pro
vides
the
best
result
in
terms
of
precision.
Based
on
the
four
performance
metrics
in
T
able
2,
Y
OLOv9
is
selected
as
the
best
o
v
erall.
Although
Y
OLOv7
and
Y
OLOv8
perform
better
than
Y
OLOv9
in
Precision,
Y
OLOv9
is
preferred
due
to
its
superior
performance
in
the
more
critical
metrics
of
mAP50
and
mAP50-95.
Figure
5
presents
the
nal
prediction
res
ults.
Figures
5(a)
and
5(b)
pro
vide
tw
o
sample
nal
predic-
tions.
From
our
proposed
system,
users
can
determine
whether
tw
o
co
ws
are
the
same
or
not
based
on
the
muzzle.
The
SSI
v
alues
of
the
inputs
in
Figures
5(a)
and
5(b)
are
0.6970
and
0
.581756,
respecti
v
ely
.
If
the
SSI
is
greater
than
0.66,
the
co
ws
are
considered
to
be
the
same.
Therefore,
the
ultimate
outputs
for
Figures
5(a)
and
5(b)
are
the
same
and
dif
ferent,
respecti
v
ely
.
These
results
indicate
that
the
proposed
algorithm
is
capable
of
making
accurate
predictions.
Moreo
v
er
,
we
ha
v
e
added
a
unique
feature
to
the
proposed
system
named
“Self-Enhancing”
using
an
LLM.
Figure
6
sho
ws
three
sample
muzzle
predictions
generated
by
the
LLM.
Finally
,
Figure
7
presents
the
performance
comparison
of
the
three
Y
OLO
v
ersions.
Figures
7(a)-7(c)
sho
w
the
performance
of
Y
OLOv7,
Y
OLOv8,
and
Y
OLOv9,
respecti
v
ely
,
based
on
Precision,
Recall,
mAP50,
and
mAP50-95.
T
able
2.
Performance
comparison
Model
Name
Precision
Recall
mAP50
mAP50-95
Y
OLO
v7
0.761
0.239
0.351
0.0816
Y
OLO
v8
0.585
0.56
0.535
0.304
Y
OLO
v9
0.391
0.916
0.559
0.307
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
4,
December
2026:
1861–1870
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
1867
(a)
(b)
Figure
5.
Sample
ultimate
prediction
(a)
same
co
ws
and
(b)
dif
ferent
co
ws
Figure
6.
Sample
prediction
using
LLM
(a)
(b)
(c)
Figure
7.
T
raining
performance
comparison
of
the
e
v
aluated
Y
OLO
models;
(a)
performance
of
the
trained
Y
OLOv7
model,
(b)
performance
of
the
trained
Y
OLOv8
model,
and
(c)
performance
of
the
trained
Y
OLOv9
model
A
muzzle-based
self-enhancing
cow
r
eco
gnition
method
.
.
.
(Rifat
Sark
er
Aoyon)
Evaluation Warning : The document was created with Spire.PDF for Python.
1868
❒
ISSN:
2252-8776
5.
CONCLUSION
In
this
research,
we
ha
v
e
demonstrated
an
insane
method
and
also
presented
the
ef
fecti
v
eness
of
the
performance
of
the
proposed
frame
w
ork
where
Y
OLO
and
LLM
w
ork
together
to
deli
v
er
appropriate
output
to
the
users.
Moreo
v
er
,
we
ha
v
e
also
compared
the
performance
of
three
types
of
Y
OLO
v
ersions
named
Y
OLOv7,
Y
OLOv8,
and
Y
OLOv9.
On
the
other
hand,
it
is
also
pro
v
ed
that
Y
OLO
is
well
capable
of
learning
with
a
little
amount
of
Roboo
w-annotated
data.
That
has
sho
wn
potential
to
implement
the
services
of
Y
OLO
and
Roboo
w
in
man
y
industry-friendly
products.
Along
with
this,
the
inte
gration
of
LLM
into
the
system
has
made
the
scope
to
implement
the
system
instantly
to
the
users
without
an
y
data,
and
that
becomes
feasible
because
of
ha
ving
a
feature
in
the
system
named
self-enhancing.
That
is
ho
w
we
accomplished
the
research
of
making
a
user
-friendly
co
w
identication
fram
e
w
ork
with
lo
w
computational
resources
and
auto-impro
ving
ability
.
In
the
future,
we
ha
v
e
the
moti
v
e
to
de
v
elop
this
frame
w
ork
by
ensuring
the
data
securit
y
of
the
users
by
implementing
distrib
uted
ledger
technology
lik
e
blockchain.
This
impro
v
ement
will
open
more
doors
to
use
this
frame
w
ork
where
data
security
is
highly
essential.
Also,
we
ha
v
e
another
objecti
v
e
to
add
of
ine
LLM
lik
e
LLaV
A
in
the
future.
FUNDING
INFORMA
TION
This
research
is
funded
by
W
oosong
Uni
v
ersity
Academic
Research
2026.
A
UTHOR
CONTRIB
UTIONS
ST
A
TEMENT
This
journal
uses
the
Contrib
utor
Roles
T
axonomy
(CRediT)
to
recognize
indi
vidual
author
contrib
utions,
reduce
authorship
disputes,
and
f
acilitate
collaboration.
Name
of
A
uthor
C
M
So
V
a
F
o
I
R
D
O
E
V
i
Su
P
Fu
Rif
at
Sark
er
Ao
yon
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Ismail
Hossain
✓
✓
✓
✓
✓
✓
✓
Jia
Uddin
✓
✓
✓
✓
✓
C
:
C
onceptualization
I
:
I
n
v
estig
ation
V
i
:
V
i
sualization
M
:
M
ethodology
R
:
R
esources
Su
:
Su
pervision
So
:
So
ftw
are
D
:
D
ata
Curation
P
:
P
roject
Administrati
on
V
a
:
V
a
lidation
O
:
Writing
-
O
riginal
Draft
Fu
:
Fu
nding
Acquisition
F
o
:
F
o
rmal
Analysis
E
:
Writing
-
Re
vie
w
&
E
diting
CONFLICT
OF
INTEREST
ST
A
TEMENT
The
authors
state
no
conict
of
interest.
D
A
T
A
A
V
AILABILITY
The
dataset
is
a
v
ailable
at
this
link:
https://github
.com/JiaUddinPhD/Co
w-Muzzle-Dataset.git.
REFERENCES
[1]
T
.
Mihret,
Mitku,
Fentahun,
and
T
.
Guadu,
“Dairy
f
arming
and
its
economic
importance
in
Ethiopia:
a
re
vie
w
,
”
W
orld
J
ournal
of
Dairy
&
F
ood
Sciences
,
v
ol.
12,
no.
1,
pp.
42–51,
2017,
doi:
10.5829/idosi.wjdfs.2017.42.51.
[2]
R
.
Grant
and
W
.
H.
Miner
,
“Economic
benets
of
impro
v
ed
co
w
comfort,
”
No
vus
International,
Inc.
,
no.
April,
pp.
1–26,
2015,
[Online].
A
v
ailable:
https://www
.dairychallenge.or
g/pdfs/2015
National/resources/No
vus
Economic
Benets
of
Impro
v
ed
Co
w
Comfort
April
2015.pdf.
[3]
J.
T
ri
v
edi
and
B.
K.
Soni,
“
A
study
on
cattle
insurance
in
villages
of
g
andhinag
ar
di
strict,
”
Business
P
er
spectives
and
Resear
c
h
,
v
ol.
2,
no.
2,
pp.
25–36,
2014,
doi:
10.1177/2278533720140204.
[4]
A.
I.
A
w
ad,
“From
classical
methods
to
animal
biometrics:
a
re
vie
w
on
cattle
identication
and
tracking,
”
Computer
s
and
Electr
onics
in
Agricultur
e
,
v
ol.
123,
pp.
423–435,
Apr
.
2016,
doi:
10.1016/j.compag.2016.03.014.
[5]
M
.-J.
Chen
and
A.
C.
Bo
vik,
“F
ast
structural
similarity
inde
x
algorithm,
”
J
our
nal
of
Real-T
ime
Ima
g
e
Pr
ocessing
,
v
ol.
6,
no.
4,
pp.
281–287,
Dec.
2011,
doi:
10.1007/s11554-010-0170-9.
[6]
E.
Hember
g,
S.
Moskal
,
and
U.-M.
O’Rei
lly
,
“Ev
olving
code
wi
th
a
lar
ge
language
model,
”
Genetic
Pr
o
gr
amming
and
Evolvable
Mac
hines
,
v
ol.
25,
no.
2,
p.
21,
Dec.
2024,
doi:
10.1007/s10710-024-09494-2.
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
4,
December
2026:
1861–1870
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Inf
&
Commun
T
echnol
ISSN:
2252-8776
❒
1869
[7]
S.
Zhou
et
al.
,
“
An
accurate
detection
model
of
takifugu
rubripes
using
an
impro
v
ed
Y
OLO-V7
netw
ork,
”
J
ournal
of
Marine
Science
and
Engineering
,
v
ol.
11,
no.
5,
p.
1051,
May
2023,
doi:
10.3390/jmse11051051.
[8]
F
.
M.
T
alaat
and
H.
ZainEldin,
“
An
impro
v
ed
re
detection
approa
ch
based
on
Y
OLO-v8
for
smart
cities,
”
Neur
al
Computing
and
Applications
,
v
ol.
35,
no.
28,
pp.
20939–20954,
Oct.
2023,
doi:
10.1007/s00521-023-08809-1.
[9]
Z.
Li,
F
.
Liu,
W
.
Y
ang,
S.
Peng,
and
J.
Zhou,
“
A
surv
e
y
of
con
v
olutional
neural
netw
orks:
analysis,
applications,
and
prospects,
”
IEEE
tr
ansactions
on
neur
al
networks
and
learning
systems
,
v
ol.
33,
no.
12,
pp.
6999–7019,
2021.
[10]
M.
A.
R.
Alif
and
M.
Hussain,
“Y
OLOv1
to
Y
OLOv10:
a
comprehensi
v
e
re
vie
w
of
Y
OLO
v
ari-
ants
and
their
application
in
the
agricultural
domain,
”
arXiv
pr
eprint
arXiv:2406.10139
,
2024.
[11]
Z.
Qi
et
al.
,
“W
ebRL:
training
LLM
web
agents
via
self-e
v
olving
online
curriculum
reinforcement
learning,
”
in
13th
International
Confer
ence
on
Learning
Repr
esentations,
ICLR
2025
,
Jan.
2025,
pp.
65879–65909.
[Online].
A
v
ailable:
http://arxi
v
.or
g/abs/2411.02337
[12]
T
.
T
.
Zin
et
al.
,
“Co
w
identication
system
us
ing
ear
tag
recognition,
”
in
2020
IEEE
2nd
Global
Confer
ence
on
Life
Sciences
and
T
ec
hnolo
gies
(LifeT
ec
h)
,
IEEE,
Mar
.
2020,
pp.
65–66,
doi:
10.1109/LifeT
ech48969.2020.1570625232.
[13]
Y
.
Lu,
X.
He,
Y
.
W
en,
and
P
.
S.
P
.
W
ang,
“
A
ne
w
co
w
identicat
ion
system
based
on
iris
analysis
and
recognition,
”
International
J
ournal
of
Biometrics
,
v
ol.
6,
no.
1,
p.
18,
2014,
doi:
10.1504/IJBM.2014.059639.
[14]
J.
Q.
Gu,
Z.
H.
W
ang,
R.
H.
Gao,
and
H.
R.
W
u,
“Co
w
beha
vior
recognition
based
on
image
analysis
and
acti
vities,
”
International
J
ournal
of
Agricultur
al
and
Biolo
gical
Engineering
,
v
ol.
10,
no.
3,
pp.
165–174,
2017,
doi:
10.3965/j.ijabe.20171003.3080.
[15]
L.
Y
ao,
Z.
Hu,
C.
Liu,
H.
Liu,
Y
.
K
uang,
and
Y
.
Gao,
“Co
w
f
ace
detection
and
recognition
based
on
automatic
feature
e
xtraction
algorithm,
”
in
Pr
oceedings
of
the
A
CM
T
uring
Celebr
ation
Confer
ence
-
China
,
Ne
w
Y
ork,
NY
,
USA:
A
CM,
May
2019,
pp.
1–5,
doi:
10.1145/3321408.3322628.
[16]
S.
Li
et
al.
,
“Indi
vidual
dairy
co
w
identication
based
on
lightweight
con
v
olutional
neural
netw
ork,
”
PLOS
ONE
,
v
ol.
16,
no.
11,
p.
e0260510,
No
v
.
2021,
doi:
10.1371/journal.pone.0260510.
[17]
T
.
T
.
Zin,
C.
N.
Ph
yo,
P
.
T
in,
H.
Hama,
and
I.
K
obayashi,
“Image
technology
based
co
w
identication
system
using
deep
learning,
”
in
Lectur
e
Notes
in
Engineering
and
Computer
Science
,
2018,
pp.
320–323.
[18]
S.
K
umar
,
S.
K.
Singh,
T
.
Dutta,
and
H.
P
.
Gupta,
“
A
f
ast
cattle
recognition
system
using
smart
de
vices,
”
in
Pr
oceedings
of
the
24th
A
CM
international
confer
ence
on
Multimedia
,
Ne
w
Y
ork,
NY
,
USA:
A
CM,
Oct.
2016,
pp.
742–743,
doi:
10.1145/2964284.2973829.
[19]
R.-W
.
Bello,
A.
Za
w
a
wi,
A.
Sufril,
D.
A.,
and
F
.
Noah,
“Image-based
indi
vidual
co
w
recognition
usi
ng
body
patterns,
”
International
J
ournal
of
Advanced
Computer
Science
and
Applications
,
v
ol.
11,
no.
3,
pp.
92–98,
2020,
doi:
10.14569/IJ
A
CSA.2020.0110311.
[20]
T
.
T
.
Zin
et
al.
,
“
Automatic
co
w
location
tracking
system
using
ear
tag
visual
anal
ysis,
”
Sensor
s
,
v
ol.
20,
no.
12,
p.
3564,
Jun.
2020,
doi:
10.3390/s20123564.
[21]
R.
Bello,
A.
Z.
H.
T
alib,
and
A.
S.
A.
Bin
Mohamed,
“Deep
learning-based
architectures
for
recognition
of
co
w
using
co
w
nose
image
pattern,
”
Gazi
Univer
sity
J
ournal
of
Science
,
v
ol.
33,
no.
3,
pp.
831–844,
Sep.
2020,
doi:
10.35378/gujs.605631.
[22]
S.
K
umar
,
S
.
T
iw
ari,
and
S.
K.
Singh,
“F
ace
recognition
for
cattle,
”
in
2015
Thir
d
International
Confer
ence
on
Ima
g
e
Information
Pr
ocessing
(ICIIP)
,
IEEE,
Dec.
2015,
pp.
65–72,
doi:
10.1109/ICIIP
.2015.7414742.
[23]
Z.
W
ang,
Z.
Hua,
Y
.
W
en,
S.
Zhang,
X.
Xu,
and
H.
Song,
“E-Y
OLO:
recognition
of
estrus
co
w
based
on
impro
v
ed
Y
OLOv8n
model,
”
Expert
Systems
with
Applications
,
v
ol.
238,
p.
122212,
Mar
.
2024,
doi:
10.1016/j.esw
a.2023.122212.
[24]
K.
Li,
D.
F
an,
H.
W
u,
and
A.
Zhao,
“
A
ne
w
dataset
for
video-based
co
w
beha
vior
recognition,
”
Scientic
Reports
,
v
ol.
14,
no.
1,
p.
18702,
Aug.
2024,
doi:
10.1038/s41598-024-65953-x.
[25]
Z.
W
eng,
F
.
Meng,
S.
Liu,
Y
.
Zhang,
Z.
Zheng,
and
C.
Gong,
“Cattle
f
ace
recognition
based
on
a
tw
o-branch
con
v
olutional
neural
netw
ork,
”
Computer
s
and
Electr
onics
in
Agricultur
e
,
v
ol.
196,
p.
106871,
May
2022,
doi:
10.1016/j.compag.2022.106871.
[26]
N
.
K
umar
,
A.
Sharma,
A.
K
umar
,
R.
Singh,
and
S.
K.
Singh,
“Cattle
v
erication
with
Y
OLO
and
cross-attention
encoder
-based
pairwise
triplet
loss,
”
Computer
s
and
Electr
onics
in
Agricultur
e
,
v
ol.
234,
p.
110223,
Jul.
2025,
doi:
10.1016/j.compag.2025.110223.
[27]
I.
Bakhshayeshi,
E.
Erf
ani,
F
.
R.
T
aghikhah,
S.
Elbourn,
A.
Beheshti,
and
M.
Asadnia,
“
An
intelligence
cattle
reidentication
system
o
v
er
transport
by
siamese
neural
netw
orks
and
Y
OLO,
”
IEEE
Internet
of
Things
J
ournal
,
v
ol.
11,
no.
2,
pp.
2351–2363,
Jan.
2024,
doi:
10.1109/JIO
T
.2023.3294944.
[28]
J.
Uddin,
“GitHub
-
JiaUddinPhD/Co
w-Muzzle-Dataset,
”
GitHub
.
[Online].
A
v
ailable:
https://github
.com/JiaUddinPhD/Co
w-
Muzzle-Dataset
[29]
I.
Bakuro
v
,
M.
Buzzelli,
R.
Schett
ini,
M.
Castelli,
and
L.
V
anneschi,
“Structural
similarity
inde
x
(SSIM)
re
visited:
a
data-dri
v
en
approach,
”
Expert
Systems
with
Applications
,
v
ol.
189,
p.
116087,
Mar
.
2022,
doi:
10.1016/j.esw
a.2021.116087.
[30]
D.
Brunet,
E.
R.
Vrscay
,
and
Z.
W
ang,
“On
the
mathematical
properties
of
the
structural
simil
arity
inde
x,
”
IEEE
T
r
ansactions
on
Ima
g
e
Pr
ocessing
,
v
ol.
21,
no.
4,
pp.
1488–1495,
Apr
.
2012,
doi:
10.1109/TIP
.2011.2173206.
[31]
OpenAI
et
al.
,
“GPT
-4o
syst
em
card,
”
arXi
v
preprint
arXi
v:2410.21276,
2024,
[Online].
A
v
ailable:
http://arxi
v
.or
g/abs/2410.21276
[32]
M.
Buckland
and
F
.
Ge
y
,
“The
relationship
between
recall
and
precision,
”
J
ournal
of
the
American
Society
for
Information
Science
,
v
ol.
45,
no.
1,
pp.
12–19,
Jan.
1994,
doi:
10.1002/(SICI)1097-4571(199401)45:1¡12::AID-ASI2¿3.0.CO;2-L.
[33]
M.
S.
M.
Sajjadi,
O.
Bousquet,
O.
Bachem,
M.
Lucic,
and
S.
Gelly
,
“
Assessing
generati
v
e
models
via
precision
and
recall,
”
Advances
in
Neur
al
Information
Pr
ocessing
Systems
,
v
ol.
2018-December
,
pp.
5228–5237,
2018.
[34]
X.
Li,
W
.
W
ang,
L.
Sun,
B.
Hu,
L.
Zhu,
and
J.
Zhang,
“Deep
learning-based
defects
detection
of
certain
aero-engine
blades
and
v
anes
with
DDSC-Y
OLOv5s,
”
Scientic
Reports
,
v
ol.
12,
no.
1,
p.
13067,
Jul.
2022,
doi:
10.1038/s41598-022-17340-7.
[35]
D.
Reis,
J.
K
upec,
J.
Hong,
and
A.
Daoudi,
“Real-time
ying
object
detection
with
Y
OLOv8,
”
arXiv
pr
eprint
arXiv:2305.09972,
2024
,
[Online].
A
v
ailable:
http://arxi
v
.or
g/abs/2305.09972
A
muzzle-based
self-enhancing
cow
r
eco
gnition
method
.
.
.
(Rifat
Sark
er
Aoyon)
Evaluation Warning : The document was created with Spire.PDF for Python.
1870
❒
ISSN:
2252-8776
BIOGRAPHIES
OF
A
UTHORS
Rifat
Sark
er
Aoy
on
recei
v
ed
the
B.S.
de
gree
in
computer
science
from
BRA
C
Uni
v
ersity
,
Dhaka,
Bangladesh,
in
2023.
He
is
currently
pursuing
a
master’
s
de
gree
in
computer
engineering
at
Chosun
Uni
v
ersity
,
Gw
angju,
South
K
orea.
He
has
authored
se
v
eral
publications
in
conference
pro-
ceedings
and
journals.
He
is
currently
w
orking
as
a
graduate
research
assistant
at
Chosun
Uni
v
ersity
,
South
K
orea.
His
research
interests
include
deep
learning,
federated
learning,
mobile
computing,
and
computer
vision.
He
can
be
contacted
at
email:
rif
at.sark
er
.ao
yon@g.bracu.ac.bd.
Ismail
Hossain
is
an
educator
,
scientist,
and
softw
are
engineer
.
He
is
a
softw
are
engineer
in
MissionSquare
Retirements,
W
ashington
DC,
USA.
He
is
also
a
lecturer
(currently
on
lea
v
e)
in
the
CSE
department
at
Brac
Uni
v
ersity
,
Bangladesh.
He
recei
v
ed
his
M.S.
in
Computer
Science
De
gree
from
Geor
ge
Mason
Uni
v
ersity
,
V
ir
ginia,
US
A,
and
completed
his
B.Sc.
in
Computer
Science
from
Brac
Uni
v
ersity
,
Bangladesh.
His
research
interest
spans
lar
ge
language
models,
machine
transla-
tions,
social
data
mining,
and
lifelong
learning.
He
can
be
contacted
at
email:
ihossai4@gmu.edu.
Jia
Uddin
is
an
associate
professor
in
the
AI
and
Big
Data
Department,
Endicott
Col
le
ge,
W
oosong
Uni
v
ersity
,
K
orea.
He
w
as
an
associate
professor
in
the
CSE
department
at
BRA
C
Uni
v
er
-
sity
,
Bangladesh.
He
recei
v
ed
a
Ph.D.
de
gree
(Computer
Engineering)
from
the
Uni
v
ersity
of
Ulsan,
South
K
orea
in
January
2015
and
an
M.Sc.
in
T
elecommunications
from
Blekinge
Institute
of
T
ech-
nology
,
Sweden,
in
2010.
He
w
as
a
member
of
the
Self-Assessment
T
eam
(SA
C)
of
CSE,
BRA
CU
in
the
HEQEP
project
funded
by
the
W
orld
Ba
nk
and
the
Uni
v
ersity
of
Grant
Commission
Bangladesh
in
2016–2017.
His
research
area
includes
f
ault
diagnosis
using
AI,
audio
and
image
processing.
He
can
be
contacted
at
email:
jia.uddin@wsu.ac.kr
.
Int
J
Inf
&
Commun
T
echnol,
V
ol.
15,
No.
4,
December
2026:
1861–1870
Evaluation Warning : The document was created with Spire.PDF for Python.