Agronomy,
Год журнала:
2023,
Номер
14(1), С. 95 - 95
Опубликована: Дек. 30, 2023
Accurate
litchi
identification
is
of
great
significance
for
orchard
yield
estimations.
Litchi
in
natural
scenes
have
large
differences
scale
and
are
occluded
by
leaves,
reducing
the
accuracy
detection
models.
Adopting
traditional
horizontal
bounding
boxes
will
introduce
a
amount
background
overlap
with
adjacent
frames,
resulting
reduced
accuracy.
Therefore,
this
study
innovatively
introduces
use
rotation
box
model
to
explore
its
capabilities
scenarios
occlusion
small
targets.
First,
dataset
on
constructed.
Secondly,
three
improvement
modules
based
YOLOv8n
proposed:
transformer
module
introduced
after
C2f
eighth
layer
backbone
network,
an
ECA
attention
added
neck
network
improve
feature
extraction
160
×
head
enhance
target
detection.
The
test
results
show
that,
compared
model,
proposed
improves
precision
rate,
recall
mAP
11.7%,
5.4%,
7.3%,
respectively.
In
addition,
four
state-of-the-art
mainstream
networks,
namely,
MobileNetv3-small,
MobileNetv3-large,
ShuffleNetv2,
GhostNet,
studied
comparison
performance
model.
article
exhibits
better
dataset,
precision,
recall,
reaching
84.6%,
68.6%,
79.4%,
This
research
can
provide
reference
estimations
complex
environments.
Scientific Reports,
Год журнала:
2025,
Номер
15(1)
Опубликована: Март 12, 2025
Urban
infrastructure,
particularly
in
ageing
cities,
faces
significant
challenges
maintaining
building
aesthetics
and
structural
integrity.
Traditional
methods
for
detecting
diseases
on
exteriors,
such
as
manual
inspections,
are
often
inefficient,
costly,
prone
to
errors,
leading
incomplete
assessments
delayed
maintenance
actions.
This
study
explores
the
application
of
advanced
deep
learning
techniques
accurately
detect
exterior
surfaces
buildings
urban
environments,
aiming
enhance
detection
efficiency
accuracy
while
providing
a
real-time
monitoring
solution
that
can
be
widely
implemented
infrastructure
health
management.
The
research
model
improves
feature
extraction
by
integrating
DenseNet
blocks
Swin-Transformer
prediction
heads,
trained
validated
using
dataset
289
high-resolution
images
collected
from
diverse
environments
China.
Data
augmentation
improved
model's
robustness
against
varying
conditions.
proposed
achieved
high
rate
84.42%,
recall
77.83%,
an
F1
score
0.81,
with
speed
55
frames
per
second.
These
metrics
demonstrate
effectiveness
identifying
complex
damage
patterns,
minute
cracks,
even
within
noisy
significantly
outperforming
traditional
methods.
highlights
potential
transform
strategies
offering
practical
ultimately
enhancing
contributing
practices
timely
interventions.
Electronics,
Год журнала:
2024,
Номер
13(15), С. 3008 - 3008
Опубликована: Июль 30, 2024
Detecting
and
recognizing
pests
are
paramount
for
ensuring
the
healthy
growth
of
crops,
maintaining
ecological
balance,
enhancing
food
production.
With
advancement
artificial
intelligence
technologies,
traditional
pest
detection
recognition
algorithms
based
on
manually
selected
features
have
gradually
been
substituted
by
deep
learning-based
algorithms.
In
this
review
paper,
we
first
introduce
primary
neural
network
architectures
evaluation
metrics
in
field
recognition.
Subsequently,
summarize
widely
used
public
datasets
Following
this,
present
various
proposed
recent
years,
providing
detailed
descriptions
each
algorithm
their
respective
performance
metrics.
Finally,
outline
challenges
that
current
encounter
propose
future
research
directions
related
Applied Sciences,
Год журнала:
2024,
Номер
14(17), С. 7695 - 7695
Опубликована: Авг. 31, 2024
Efficient
diagnosis
of
apple
diseases
and
pests
is
crucial
to
the
healthy
development
industry.
However,
existing
single-source
image-based
classification
methods
have
limitations
due
constraints
input
image
information,
resulting
in
low
accuracy
poor
stability.
Therefore,
a
method
for
disease
pest
areas
based
on
multi-source
fusion
proposed
this
paper.
Firstly,
RGB
images
multispectral
are
obtained
using
drones
construct
an
canopy
dataset.
Secondly,
vegetation
index
selection
saliency
attention
proposed,
which
uses
multi-label
ReliefF
feature
algorithm
obtain
importance
scores
indices,
enabling
automatic
indices.
Finally,
area
model
named
AMMFNet
constructed,
effectively
combines
advantages
images,
performs
data-level
data,
channel
mechanisms
exploit
complementary
aspects
between
data.
The
experimental
results
demonstrated
that
achieves
significant
subset
92.92%,
sample
85.43%,
F1
value
86.21%
dataset,
representing
improvements
8.93%
10.9%
compared
prediction
only
or
images.
also
proved
can
provide
technical
support
coarse-grained
positioning
orchards
has
good
application
potential
planting
