IEEE Access,
Journal Year:
2023,
Volume and Issue:
11, P. 69295 - 69309
Published: Jan. 1, 2023
Colorectal
polyps
is
a
prevalent
medical
condition
that
could
lead
to
colorectal
cancer,
leading
cause
of
cancer-related
mortality
globally,
if
left
undiagnosed.
Colonoscopy
remains
the
gold
standard
for
detection
and
diagnosis
neoplasia;
however,
significant
proportion
neoplastic
lesions
are
missed
during
routine
examinations,
particularly
diminutive
flat
lesions.
Deep
learning
techniques
have
been
employed
improve
polyp
rates
in
colonoscopy
images
proven
successful
reducing
miss
rate.
However,
accurate
segmentation
small
major
challenge
existing
models
as
they
struggle
differentiate
polypoid
non-polypoid
regions
apart.
To
address
this
issue,
we
present
an
enhanced
version
Multi-Scale
Attention
Network
(MA-NET)
incorporates
modified
Mix-ViT
transformer
feature
extractor.
The
facilitates
ultra-fine-grained
visual
categorization
accuracy
regions.
Additionally,
introduce
pre-processing
layer
performs
histogram
equalization
on
input
CIEL*A*B*
color
space
enhance
their
features.
Our
model
was
trained
combined
dataset
comprising
Kvasir-SEG
CVC-ColonDB
cross-validated
CVC-ClinicDB
ETIS-LaribDB.
proposed
method
demonstrates
superior
performance
compared
methods,
polyps.
Applied Sciences,
Journal Year:
2025,
Volume and Issue:
15(5), P. 2789 - 2789
Published: March 5, 2025
Unmanned
aerial
vehicle
(UAV)
imagery
often
suffers
from
significant
object
scale
variations,
high
target
density,
and
varying
distances
due
to
shooting
conditions
environmental
factors,
leading
reduced
robustness
low
detection
accuracy
in
conventional
models.
To
address
these
issues,
this
study
adopts
DGBL-YOLOv8s,
an
improved
model
tailored
for
UAV
perspectives
based
on
YOLOv8s.
First,
a
Dilated
Wide
Residual
(DWR)
module
is
introduced
replace
the
C2f
backbone
network
of
YOLOv8,
enhancing
model’s
capability
capture
fine-grained
features
contextual
information.
Second,
neck
structure
redesigned
by
incorporating
Global-to-Local
Spatial
Aggregation
(GLSA)
combined
with
Bidirectional
Feature
Pyramid
Network
(BiFPN),
which
strengthens
feature
fusion.
Third,
lightweight
shared
convolution
head
proposed,
batch
normalization
techniques.
Additionally,
further
improve
small
detection,
dedicated
small-object
introduced.
Results
experiments
VisDrone
dataset
reveal
that
DGBL-YOLOv8s
enhances
8.5%
relative
baseline
model,
alongside
34.8%
reduction
parameter
count.
The
overall
performance
exceeds
most
current
models,
confirms
advantages
proposed
improvement.
IEEE Access,
Journal Year:
2023,
Volume and Issue:
11, P. 69295 - 69309
Published: Jan. 1, 2023
Colorectal
polyps
is
a
prevalent
medical
condition
that
could
lead
to
colorectal
cancer,
leading
cause
of
cancer-related
mortality
globally,
if
left
undiagnosed.
Colonoscopy
remains
the
gold
standard
for
detection
and
diagnosis
neoplasia;
however,
significant
proportion
neoplastic
lesions
are
missed
during
routine
examinations,
particularly
diminutive
flat
lesions.
Deep
learning
techniques
have
been
employed
improve
polyp
rates
in
colonoscopy
images
proven
successful
reducing
miss
rate.
However,
accurate
segmentation
small
major
challenge
existing
models
as
they
struggle
differentiate
polypoid
non-polypoid
regions
apart.
To
address
this
issue,
we
present
an
enhanced
version
Multi-Scale
Attention
Network
(MA-NET)
incorporates
modified
Mix-ViT
transformer
feature
extractor.
The
facilitates
ultra-fine-grained
visual
categorization
accuracy
regions.
Additionally,
introduce
pre-processing
layer
performs
histogram
equalization
on
input
CIEL*A*B*
color
space
enhance
their
features.
Our
model
was
trained
combined
dataset
comprising
Kvasir-SEG
CVC-ColonDB
cross-validated
CVC-ClinicDB
ETIS-LaribDB.
proposed
method
demonstrates
superior
performance
compared
methods,
polyps.
