Algorithms,
Год журнала:
2024,
Номер
17(12), С. 591 - 591
Опубликована: Дек. 21, 2024
Building
on
a
previously
developed
partially
synthetic
data
generation
algorithm
utilizing
visualization
techniques,
this
study
extends
the
novel
to
generate
fully
tabular
healthcare
data.
In
enhanced
form,
serves
as
an
alternative
conventional
methods
based
Generative
Adversarial
Networks
(GANs)
or
Variational
Autoencoders
(VAEs).
By
iteratively
applying
original
methodology,
adapted
employs
UMAP
(Uniform
Manifold
Approximation
and
Projection),
dimensionality
reduction
technique,
validate
generated
samples
through
low-dimensional
clustering.
This
approach
has
been
successfully
applied
three
domains:
prostate
cancer,
breast
cardiovascular
disease.
The
have
rigorously
evaluated
for
fidelity
utility.
Results
show
that
UMAP-based
outperforms
GAN-
VAE-based
across
different
scenarios.
assessments,
it
achieved
smaller
maximum
distances
between
cumulative
distribution
functions
of
real
attributes.
utility
evaluations,
datasets
machine
learning
model
performance,
particularly
in
classification
tasks.
conclusion,
method
represents
robust
solution
generating
secure,
high-quality
data,
effectively
addressing
scarcity
challenges.
Drones,
Год журнала:
2024,
Номер
8(4), С. 161 - 161
Опубликована: Апрель 19, 2024
A
UAV
infrared
target
detection
model
ITD-YOLOv8
based
on
YOLOv8
is
proposed
to
address
the
issues
of
missed
and
false
detections
caused
by
complex
ground
background
uneven
scale
in
aerial
image
detection,
as
well
high
computational
complexity.
Firstly,
an
improved
backbone
feature
extraction
network
designed
lightweight
GhostHGNetV2.
It
can
effectively
capture
information
at
different
scales,
improving
accuracy
environments
while
remaining
lightweight.
Secondly,
VoVGSCSP
improves
perceptual
abilities
referencing
global
contextual
multiscale
features
enhance
neck
structure.
At
same
time,
a
convolutional
operation
called
AXConv
introduced
replace
regular
module.
Replacing
traditional
fixed-size
convolution
kernels
with
sizes
reduces
complexity
model.
Then,
further
optimize
reduce
during
object
CoordAtt
attention
mechanism
weight
channel
dimensions
map,
allowing
pay
more
important
information,
thereby
robustness
detection.
Finally,
implementation
XIoU
loss
function
for
boundary
boxes
enhances
precision
localization.
The
experimental
findings
demonstrate
that
ITD-YOLOv8,
comparison
YOLOv8n,
rate
detecting
multi-scale
small
targets
backgrounds.
Additionally,
it
achieves
41.9%
reduction
parameters
25.9%
decrease
floating-point
operations.
Moreover,
mean
(mAP)
attains
impressive
93.5%,
confirming
model’s
applicability
unmanned
vehicles
(UAVs).
Drones,
Год журнала:
2024,
Номер
8(9), С. 495 - 495
Опубликована: Сен. 18, 2024
A
lightweight
infrared
target
detection
model,
G-YOLO,
based
on
an
unmanned
aerial
vehicle
(UAV)
is
proposed
to
address
the
issues
of
low
accuracy
in
UAV
images
complex
ground
scenarios
and
large
network
models
that
are
difficult
apply
mobile
or
embedded
platforms.
Firstly,
YOLOv8
backbone
feature
extraction
improved
designed
network,
GhostBottleneckV2,
remaining
part
adopts
depth-separable
convolution,
DWConv,
replace
standard
which
effectively
retains
effect
model
while
greatly
reducing
number
parameters
calculations.
Secondly,
neck
structure
by
ODConv
module,
adaptive
convolutional
adaptively
adjust
kernel
size
step
size,
allows
for
more
effective
targets
at
different
scales.
At
same
time,
further
optimized
using
attention
mechanism,
SEAttention,
improve
model’s
ability
learn
global
information
input
maps,
then
applied
each
channel
map
enhance
useful
a
specific
performance.
Finally,
introduction
SlideLoss
loss
function
enables
calculate
differences
between
predicted
actual
truth
bounding
boxes
during
training
process,
these
efficiency
object
detection.
The
experimental
results
show
compared
with
YOLOv8n,
G-YOLO
reduces
missed
false
rates
small
backgrounds.
reduced
74.2%,
computational
floats
54.3%,
FPS
71,
improves
average
(mAP)
reaches
91.4%,
verifies
validity
UAV-based
Furthermore,
556,
it
will
be
suitable
wider
task
such
as
targets,
long-distance
other
scenes.
PeerJ Computer Science,
Год журнала:
2025,
Номер
11, С. e2764 - e2764
Опубликована: Май 6, 2025
Thanks
to
the
presence
of
artificial
intelligence
methods,
diagnosis
patients
can
be
done
quickly
and
accurately.
This
article
introduces
a
new
diagnostic
system
(DS)
that
includes
three
main
layers
called
rejection
layer
(RL),
selection
(SL),
(DL)
accurately
diagnose
cases
suffering
from
various
diseases.
In
RL,
outliers
removed
using
genetic
algorithm
(GA).
At
same
time,
best
features
selected
by
feature
method
hybrid
approach
(HFSA)
in
SL.
next
step,
filtered
data
is
passed
naive
Bayes
(NB)
classifier
DL
give
accurate
diagnoses.
this
work,
contribution
represented
introducing
HFSA
as
composed
two
stages;
fast
stage
(FS)
(AS).
FS,
chi-square,
filtering
methodology,
applied
select
while
Hybrid
Optimization
Algorithm
(HOA),
wrapper
AS
features.
It
concluded
better
than
other
methods
based
on
experimental
results
because
enable
different
classifiers
NB,
K-nearest
neighbors
(KNN),
neural
network
(ANN)
provide
maximum
accuracy,
precision,
recall
values
minimum
error
value.
Additionally,
proved
DS,
including
GA
an
outlier
method,
selection,
NB
mode,
outperformed
models.
PLoS ONE,
Год журнала:
2025,
Номер
20(5), С. e0322738 - e0322738
Опубликована: Май 30, 2025
Outlier
detection
is
essential
for
identifying
unusual
patterns
or
observations
that
significantly
deviate
from
the
normal
behavior
of
a
dataset.
With
rapid
growth
data
science,
prevalence
anomalies
and
outliers
has
increased,
which
can
disrupt
system
modeling
parameter
estimation,
leading
to
inaccurate
results.
Recently,
deep
learning-based
outlier
methods
have
gained
significant
attention,
but
their
performance
often
limited
by
challenges
in
selection
nearest
neighbor
search.
To
overcome
these
limitations,
we
propose
three-stage
Efficient
Detection
Approach
(named
EODA),
not
only
detects
with
high
accuracy
also
emphasizes
dataset
characteristics.
In
first
stage,
apply
feature
algorithm
based
on
Boruta
method
Random
Forest
reduce
size
selecting
most
relevant
attributes
calculating
highest
Z-score
shadow
features.
second
improve
K-nearest
neighbors
enhance
identification
clustering
phase.
Finally,
third
stage
efficiently
identifies
within
clustered
datasets.
We
evaluate
proposed
EODA
across
eight
UCI
machine-learning
repository
The
results
demonstrate
effectiveness
our
approach,
achieving
Precision
63.07%,
Recall
82.49%,
an
F1-Score
64.53%,
outperforming
existing
techniques
field.
Drones,
Год журнала:
2024,
Номер
8(9), С. 479 - 479
Опубликована: Сен. 12, 2024
Deploying
target
detection
models
on
edge
devices
such
as
UAVs
is
challenging
due
to
their
limited
size
and
computational
capacity,
while
typically
require
significant
resources.
To
address
this
issue,
study
proposes
a
lightweight
real-time
infrared
object
model
named
LRI-YOLO
(Lightweight
Real-time
Infrared
YOLO),
which
based
YOLOv8n.
The
improves
the
C2f
module’s
Bottleneck
structure
by
integrating
Partial
Convolution
(PConv)
with
Pointwise
(PWConv),
achieving
more
design.
Furthermore,
during
feature
fusion
stage,
original
downsampling
ordinary
convolution
replaced
combination
of
max
pooling
regular
convolution.
This
modification
retains
map
information.
model’s
further
optimized
redesigning
decoupled
head
Group
(GConv)
instead
convolution,
significantly
enhancing
speed.
Additionally,
BCELoss
EMASlideLoss,
newly
developed
classification
loss
function
introduced
in
study.
allows
focus
hard
samples,
thereby
improving
its
capability.
Compared
YOLOv8n
algorithm,
lightweight,
parameters
reduced
46.7%
floating-point
operations
(FLOPs)
53.1%.
Moreover,
mean
average
precision
(mAP)
reached
94.1%.
Notably,
moderate
power
that
only
have
Central
Processing
Unit
(CPU),
speed
42
frames
per
second
(FPS),
surpassing
most
mainstream
models.
indicates
offers
novel
solution
for
drones.
Algorithms,
Год журнала:
2024,
Номер
17(4), С. 160 - 160
Опубликована: Апрель 15, 2024
The
pressing
need
for
sustainable
development
solutions
necessitates
innovative
data-driven
tools.
Machine
learning
(ML)
offers
significant
potential,
but
faces
challenges
in
centralized
approaches,
particularly
concerning
data
privacy
and
resource
constraints
geographically
dispersed
settings.
Federated
(FL)
emerges
as
a
transformative
paradigm
by
decentralizing
ML
training
to
edge
devices.
However,
communication
bottlenecks
hinder
its
scalability
sustainability.
This
paper
introduces
an
FL
framework
that
enhances
efficiency.
proposed
addresses
the
bottleneck
harnessing
power
of
Lemurs
optimizer
(LO),
nature-inspired
metaheuristic
algorithm.
Inspired
cooperative
foraging
behavior
lemurs,
LO
strategically
selects
most
relevant
model
updates
communication,
significantly
reducing
overhead.
was
rigorously
evaluated
on
CIFAR-10,
MNIST,
rice
leaf
disease,
waste
recycling
plant
datasets
representing
various
areas
development.
Experimental
results
demonstrate
reduces
overhead
over
15%
average
compared
baseline
while
maintaining
high
accuracy.
breakthrough
extends
applicability
resource-constrained
environments,
paving
way
more
scalable
real-world
initiatives.