CV-YOLOv10-AR-M: Foreign Object Detection in Pu-Erh Tea Based on Five-Fold Cross-Validation
Foods,
Год журнала:
2025,
Номер
14(10), С. 1680 - 1680
Опубликована: Май 9, 2025
To
address
the
problem
of
detecting
foreign
bodies
in
Pu-erh
tea,
this
study
proposes
an
intelligent
detection
method
based
on
improved
YOLOv10
network.
By
introducing
MPDIoU
loss
function,
network
is
optimized
to
effectively
enhance
positioning
accuracy
model
complex
background
and
improve
small
target
objects.
Using
AssemFormer
optimize
structure,
network’s
ability
perceive
objects
its
process
global
information
are
improved.
Rectangular
Self-Calibrated
Module,
prediction
bounding
box
optimized,
further
improving
classification
target-positioning
abilities
scenes.
The
results
showed
that
Box,
Cls,
Dfl
functions
CV-YOLOv10-AR-M
One-to-Many
Head
task
were,
respectively,
14.60%,
19.74%,
20.15%
lower
than
those
In
One-to-One
task,
they
decreased
by
10.42%,
29.11%,
20.15%,
respectively.
Compared
with
original
network,
accuracy,
recall
rate,
mAP
were
increased
5.35%,
11.72%
8.32%,
improves
model’s
attention
sizes,
backgrounds,
detailed
information,
providing
effective
technical
support
for
quality
control
agricultural
field.
Язык: Английский
MAS-YOLO: A Lightweight Detection Algorithm for PCB Defect Detection Based on Improved YOLOv12
Applied Sciences,
Год журнала:
2025,
Номер
15(11), С. 6238 - 6238
Опубликована: Июнь 1, 2025
As
the
performance
requirements
for
printed
circuit
boards
(PCBs)
in
electronic
devices
continue
to
increase,
reliable
defect
detection
during
PCB
manufacturing
is
vital.
However,
due
small
size,
complex
categories,
and
subtle
differences
features,
traditional
methods
are
limited
accuracy
robustness.
To
overcome
these
challenges,
this
paper
proposes
MAS-YOLO,
a
lightweight
algorithm
based
on
improved
YOLOv12
architecture.
In
Backbone,
Median-enhanced
Channel
Spatial
Attention
Block
(MECS)
expands
receptive
field
through
median
enhancement
depthwise
convolution
generate
attention
maps
that
effectively
capture
features.
Neck,
an
Adaptive
Hierarchical
Feature
Integration
Network
(AHFIN)
adaptively
fuses
multi-scale
features
weighted
integration,
enhancing
feature
utilization
focus
regions.
Moreover,
original
loss
function
replaced
with
Slide
Alignment
Loss
(SAL)
improve
bounding
box
localization
detect
types.
Experimental
results
demonstrate
MAS-YOLO
significantly
improves
mean
average
precision
(mAP)
frames
per
second
(FPS)
compared
YOLOv12,
fulfilling
real-time
industrial
requirements.
Язык: Английский