Sensors,
Год журнала:
2024,
Номер
24(18), С. 5910 - 5910
Опубликована: Сен. 12, 2024
Aiming
at
the
problems
of
a
large
volume,
slow
processing
speed,
and
difficult
deployment
in
edge
terminal,
this
paper
proposes
lightweight
insulator
detection
algorithm
based
on
an
improved
SSD.
Firstly,
original
feature
extraction
network
VGG-16
is
replaced
by
Ghost
Module
to
initially
achieve
model.
A
Feature
Pyramid
structure
Network
(FPN+PAN)
are
integrated
into
Neck
part
Simplified
Spatial
Pooling
Fast
(SimSPPF)
module
introduced
realize
integration
local
features
global
features.
Secondly,
multiple
Channel
Squeeze-and-Excitation
(scSE)
attention
mechanisms
make
model
pay
more
channels
containing
important
information.
The
six
heads
reduced
four
improve
inference
speed
network.
In
order
recognition
performance
occluded
overlapping
targets,
DIoU-NMS
was
used
replace
non-maximum
suppression
(NMS).
Furthermore,
channel
pruning
strategy
reduce
unimportant
weight
matrix
model,
knowledge
distillation
fine-adjust
after
pruning,
so
as
ensure
accuracy.
experimental
results
show
that
parameter
number
proposed
from
26.15
M
0.61
M,
computational
load
118.95
G
1.49
G,
mAP
increased
96.8%
98%.
Compared
with
other
models,
not
only
guarantees
accuracy
algorithm,
but
also
greatly
reduces
which
provides
support
for
realization
visible
light
target
intelligence.
Agriculture,
Год журнала:
2025,
Номер
15(7), С. 713 - 713
Опубликована: Март 27, 2025
In
a
natural
environment,
due
to
the
small
size
of
caterpillar
fungus,
its
indistinct
features,
similar
color
surrounding
weeds
and
background,
overlapping
instances
identifying
fungus
poses
significant
challenges.
To
address
these
issues,
this
paper
proposes
new
MRAA
network,
which
consists
feature
fusion
pyramid
network
(MRFPN)
backbone
N-CSPDarknet53.
MRFPN
is
used
solve
problem
weak
features.
N-CSPDarknet53,
Da-Conv
module
proposed
background
interference
problems
in
shallow
maps.
The
significantly
improves
accuracy,
achieving
an
accuracy
rate
0.202
APS
for
small-target
recognition,
represents
12%
increase
compared
baseline
0.180
APS.
Additionally,
model
(9.88
M),
making
it
lightweight.
It
easy
deploy
embedded
devices,
greatly
promotes
development
application
identification.
Research Square (Research Square),
Год журнала:
2025,
Номер
unknown
Опубликована: Апрель 7, 2025
Abstract
Pick-and-place
robots
play
a
crucial
role
in
industrial
automation,
helping
to
lower
labor
costs,
minimize
errors,
and
improve
production
efficiency.
Many
image
processing
methods
have
been
proposed
facilitate
the
pick-and-place
operation.
However,
performance
of
these
is
sensitive
lighting
conditions,
presence
occlusions,
variations
object
appearance.
Although
many
challenges
can
be
overcome
through
use
deep
learning
methods,
direct
comparison
coupled
with
an
analysis
different
picking
strategies,
lacking.
The
present
study
addresses
this
gap
by
conducting
simulation-based
evaluation
accuracy
time
ORB
algorithm
YOLOv8
model
for
recognition.
effects
two
strategies
(FIFO
Euclidean
Distance)
on
system
throughput
are
also
explored.
simulation
results
show
that
achieves
higher
(98%)
significantly
faster
(138
ms)
than
(97.33%
715.24
ms
time).
Additionally,
FIFO
strategy
improves
productivity
13%
compared
Distance
strategy.
Overall,
findings
provide
valuable
insights
into
optimizing
robotic
operations
automation
settings.
Diagnostics,
Год журнала:
2024,
Номер
14(19), С. 2234 - 2234
Опубликована: Окт. 7, 2024
Background/Objectives:
Lung
and
cardiovascular
diseases
are
leading
causes
of
mortality
worldwide,
yet
early
detection
remains
challenging
due
to
the
subtle
symptoms.
Digital
clubbing,
characterized
by
bulbous
enlargement
fingertips,
serves
as
an
indicator
these
diseases.
This
study
aims
develop
automated
system
for
detecting
digital
clubbing
using
deep-learning
models
real-time
monitoring
intervention.
Methods:
The
proposed
utilizes
YOLOv8
model
object
U-Net
image
segmentation,
integrated
with
ESP32-CAM
development
board
capture
analyze
finger
images.
severity
is
determined
a
custom
algorithm
based
on
Lovibond
angle
theory,
categorizing
condition
into
normal,
mild,
moderate,
severe.
was
evaluated
1768
images
achieved
cloud-based
processing
capabilities.
Results:
demonstrated
high
accuracy
(98.34%)
in
precision
(98.22%),
sensitivity
(99.48%),
specificity
(98.22%).
Cloud-based
slightly
lower
but
robust
results,
96.38%.
average
time
0.15
s
per
image,
showcasing
its
potential.
Conclusions:
provides
scalable
cost-effective
solution
enabling
timely
intervention
lung
Its
capabilities
make
it
suitable
both
clinical
home-based
health
monitoring.
