IEEE Access,
Год журнала:
2024,
Номер
12, С. 109367 - 109379
Опубликована: Янв. 1, 2024
The
attention
enhancement
YOLO
printed
circuit
board
(PCB)
defect
detection
algorithm
AE-YOLO,
which
improves
YOLOv8,
is
proposed
to
improve
the
current
slow
speed
of
PCB
problems,
such
as
high
missed
or
false
rates
and
low
accuracy.
First,
in
backbone
network,
CoT
Net
used
instead
original
feature
extraction
network
reduce
number
parameters
model
its
while
maintaining
accuracy
much
possible.
Then,
SPPFS
module
last
layer
enhance
model's
ability
extract
global
information,
fuse
features,
use
rich
primary
semantic
information
pave
way
for
subsequent
classification
positioning.
Finally,
CC3
perceive
high-level
help
decoupled
head
better
perform
target
prediction
positioning,
comprehensiveness
model,
provide
with
continuous
performance
improvements.
Compared
YOLOv8
AE-YOLO
compresses
by
16%,
increases
2.9%,
recall
rate
3.3%.
This
provides
a
more
efficient
method
detection.
Sensors,
Год журнала:
2024,
Номер
24(17), С. 5837 - 5837
Опубликована: Сен. 8, 2024
As
semiconductor
chip
manufacturing
technology
advances,
structures
are
becoming
more
complex,
leading
to
an
increased
likelihood
of
void
defects
in
the
solder
layer
during
packaging.
However,
identifying
packaged
chips
remains
a
significant
challenge
due
complex
background,
varying
defect
sizes
and
shapes,
blurred
boundaries
between
voids
their
surroundings.
To
address
these
challenges,
we
present
deep-learning-based
framework
for
segmentation
The
consists
two
main
components:
region
extraction
method
network.
includes
lightweight
network
rotation
correction
algorithm
that
eliminates
background
noise
accurately
captures
chip.
is
designed
efficient
accurate
segmentation.
cope
with
variability
shapes
sizes,
propose
Mamba
model-based
encoder
uses
visual
state
space
module
multi-scale
information
extraction.
In
addition,
interactive
dual-stream
decoder
feature
correlation
cross
gate
fuse
streams’
features
improve
produce
maps.
effectiveness
evaluated
through
quantitative
qualitative
experiments
on
our
custom
X-ray
dataset.
Furthermore,
proposed
packaging
has
been
applied
real
factory
inspection
line,
achieving
accuracy
93.3%
qualification.
Scientific Reports,
Год журнала:
2025,
Номер
15(1)
Опубликована: Янв. 17, 2025
The
existing
UAV
inspection
images
are
faced
with
many
challenges
for
insulator
defect
recognition.
A
new
multi-resolution
Context
Cluster
CenterNet++
model
is
proposed.
First,
this
paper
proposes
the
method
to
solve
problem
of
low
recognition
accuracy
caused
by
non-uniform
distribution
targets.
cluster
region
used
identify
and
predict
location
target,
improved
loss
function
modify
center.
Secondly,
uses
deformable
convolution
operator
(DCNv2)
combined
path
aggregation
network
(PAN)
carry
out
operation
on
image,
accurately
predicts
regression
box
key
point
triplet
(KP),
so
as
improve
accurate
positioning
target
position
any
shape
scale.
sensitivity
scale
change
deformation
reduced,
improved.
Then,
Bhattacharyya
distance
calculate
prediction
points
center
offset
loss,
significantly
same
in
different
frames.
Finally,
experiments
carried
MS-COCO
dataset
National
Grid
standardized
image
dataset.
Our
code
at
https://github.com/mengbonannan88/CC-CenterNet
.
Scientific Reports,
Год журнала:
2025,
Номер
15(1)
Опубликована: Янв. 20, 2025
Accurate
detection
of
surface
defects
on
strip
steel
is
essential
for
ensuring
product
quality.
Existing
deep
learning
based
detectors
typically
strive
to
iteratively
refine
and
integrate
the
coarse
outputs
backbone
network,
enhancing
models'
ability
express
defect
characteristics.
Attention
mechanisms
including
spatial
attention,
channel
attention
self-attention
are
among
most
prevalent
techniques
feature
extraction
fusion.
This
paper
introduces
an
innovative
triple-attention
mechanism
(TA),
characterized
by
interrelated
complementary
interactions,
that
concurrently
refines
integrates
maps
from
three
distinct
perspectives,
thereby
features'
capacity
representation.
The
idea
following
observation:
given
a
three-dimensional
map,
we
can
examine
map
different
yet
two-dimensional
planar
perspectives:
channel-width,
channel-height,
width-height
perspectives.
Based
TA,
novel
detector,
called
TADet,
proposed,
which
encoder-decoder
network:
decoder
uses
proposed
TA
refines/fuses
multiscale
rough
features
generated
encoder
(backbone
network)
perspectives
(branches)
then
purified
branches.
Extensive
experimental
results
show
TADet
superior
state-of-the-art
methods
in
terms
mean
absolute
error,
S-measure,
E-measure
F-measure,
confirming
effectiveness
robustness
TADet.
Our
code
available
at
https://github.com/hpguo1982/TADet
.
