Remote Sensing,
Год журнала:
2023,
Номер
15(11), С. 2784 - 2784
Опубликована: Май 26, 2023
Traditional
image
fusion
techniques
generally
use
symmetrical
methods
to
extract
features
from
different
sources
of
images.
However,
these
conventional
approaches
do
not
resolve
the
information
domain
discrepancy
multiple
sources,
resulting
in
incompleteness
fusion.
To
solve
problem,
we
propose
an
asymmetric
decomposition
method.
Firstly,
abundance
discrimination
method
is
used
sort
images
into
detailed
and
coarse
categories.
Then,
are
proposed
at
scales.
Next,
strategies
adopted
for
scale
features,
including
sum
fusion,
variance-based
transformation,
integrated
energy-based
Finally,
result
obtained
through
summation,
retaining
vital
both
Eight
metrics
two
datasets
containing
registered
visible,
ISAR,
infrared
were
evaluate
performance
The
experimental
results
demonstrate
that
could
preserve
more
details
than
symmetric
one,
performed
better
objective
subjective
evaluations
compared
with
fifteen
state-of-the-art
methods.
These
findings
can
inspire
researchers
consider
a
new
framework
adapt
differences
richness
images,
promote
development
technology.
Scientific Reports,
Год журнала:
2024,
Номер
14(1)
Опубликована: Авг. 1, 2024
Recently,
artificial
intelligence
(AI)-generated
resources
have
gained
popularity
because
of
their
high
effectiveness
and
reliability
in
terms
output
capacity
to
be
customized
broadened,
especially
image
generation.
Traditional
Chinese
paintings
(TCPs)
are
incomplete
color
contrast
is
insufficient,
object
reality
minimal.
However,
combining
AI
painting
(AIP)
with
TCP
remains
inadequate
uncertain
features
such
as
patterns,
styles,
color.
Hence,
an
algorithm
named
variational
fusion-based
fuzzy
accelerated
(VF2AP)
has
been
proposed
resolve
this
challenge.
Initially,
the
collected
data
source
applied
for
preprocessing
convert
it
into
a
grayscale
image.
Then,
feature
extraction
process
performed
via
fuzzy-based
local
binary
pattern
(FLBP)
brushstroke
patterns
enhance
fusion
intelligent
logic
optimize
textures
noisy
Second,
extracted
used
inputs
autoencoder
(VAE),
which
avoid
latent
space
irregularities
reconstructed
by
maintaining
minimum
reconstruction
loss.
Third,
inference
rules
variation
original
images.
Fourth,
feedback
mechanism
designed
evaluation
metrics
area
under
curve-receiver
operating
characteristic
(AUC-ROC)
analysis,
mean
square
error
(MSE),
structural
similarity
index
(SSIM),
Kullback‒Leibler
(KL)
divergence
viewer's
understanding
fused
Scientific Reports,
Год журнала:
2025,
Номер
15(1)
Опубликована: Май 15, 2025
Abstract
This
study
is
based
on
the
YOLOv7
object
detection
framework
and
conducts
comparative
experiments
early
fusion,
halfway
late
fusion
for
multispectral
pedestrian
tasks.
Traditional
tasks
typically
use
image
data
from
a
single
sensor
or
modality.
However,
in
field
of
remote
sensing,
fusing
multi-source
crucial
improving
performance.
aims
to
explore
impact
different
strategies
performance
identify
most
suitable
approach
data.
Firstly,
we
implemented
by
merging
with
visible
light
at
network’s
input
layer.
Next,
were
conducted,
middle
layers.
Finally,
performed
high
A
comprehensive
comparison
experimental
results
various
reveals
that
strategy
exhibits
outstanding
tasks,
achieving
accuracy
relatively
fast
speed.
International Journal of Computational Intelligence Systems,
Год журнала:
2024,
Номер
17(1)
Опубликована: Окт. 23, 2024
Infrared
images
of
sensitive
targets
are
difficult
to
obtain
and
cannot
meet
the
design
training
needs
target
detection
tracking
algorithms
for
mobile
platforms
such
as
aircraft.
This
paper
proposes
an
image
translation
algorithm
TransImg,
which
can
achieve
visible
light
infrared
domain
enrich
dataset.
First,
designed
a
generator
structure
consisting
deep
residual
connected
encoder
region
perception
feature
fusion
module
enhance
learning,
thereby
avoiding
issues
generating
with
insufficient
details
in
transfer
task.
Afterward,
multi-scale
discriminator
composite
loss
function
were
further
improve
effect.
Finally,
automatic
mixed-precision
strategy
was
overall
migration
architecture
accelerate
generation
images.
Experiments
have
shown
that
TransImg
has
good
accuracy,
generated
by
richer
texture
details,
faster
speed,
lower
video
memory
consumption,
performance
exceeds
mainstream
traditional
algorithm,
requirements
IET Image Processing,
Год журнала:
2024,
Номер
18(10), С. 2774 - 2787
Опубликована: Май 23, 2024
Abstract
To
effectively
enhance
the
ability
to
acquire
information
by
making
full
use
of
complementary
features
infrared
and
visible
images,
widely
used
image
fusion
algorithm
is
faced
with
challenges
such
as
loss
blurring.
In
response
this
issue,
authors
propose
a
dual‐branch
deep
hierarchical
network
(ADF‐Net)
guided
an
attention
mechanism.
Initially,
convolution
module
extracts
shallow
image.
Subsequently,
decomposition
feature
extractor
introduced,
where
in
transformer
encoder
block
(TEB)
employs
remote
process
low‐frequency
global
features,
while
CNN
(CEB)
high‐frequency
local
information.
Ultimately,
layer
based
on
TEB
CEB
produce
fused
through
encoder.
Multiple
experiments
demonstrate
that
ADF‐Net
excels
various
aspects
utilizing
two‐stage
training
appropriate
function
for
testing.
Most
of
the
existing
image
fusion
models
adopt
a
global
strategy,
which
usually
reduces
contrast
infrared
images.
In
this
paper,
we
propose
an
approach
driven
by
non-global
fuzzy
pre-enhancement
framework
(FusionFRFCM),
is
more
suitable
for
structure
(IR)
image.
A
equalization
algorithm
based
on
Fourth-order
Partial
Differential
Equation
(FPDE)
proposed,
used
to
enhance
background
region.
Due
differences
between
IR
and
visible
(VIS)
images,
hybrid
strategy
Expectation
Maximization
(EM)
Principal
Component
Analysis
(PCA)
designed.
Compared
with
other
state-of-the-art
methods,
experimental
results
show
that
proposed
has
better
performance
in
both
qualitative
quantitative
results.
addition,
verify
effectiveness
our
FusionFRFCM
practical
application,
embedded
into
RGBT
target
tracking
task
under
VOT-RGBT2019
OTCBVS
datasets.
Through
comparative
experiments,
it
found
can
easily
be
integrated
improve
accuracy
many
scenes.
Engineering Science and Technology an International Journal,
Год журнала:
2024,
Номер
54, С. 101727 - 101727
Опубликована: Июнь 1, 2024
Multimodal
medical
image
fusion
plays
an
important
role
in
clinical
applications.
However,
gradient
features
and
intensity
are
not
extracted
inadequately
methods.
To
solve
the
above
problems,
this
paper
proposes
a
Gradient-Intensity
oriented
Automatic
Encode-Decode
multimodal
lung
tumor
model
(GIAE-Net),
there
two
parallel
branches
network,
one
is
branch,
another
branch.
The
main
idea
of
proposed
network
as
follows:Firstly,
attention
module
(GAM)
designed
to
enhance
description
ability
fine-grained
spatial
by
using
operators,
so
that
can
retain
more
edge
details.
Secondly,
(IAM)
constructed
enable
learn
features,
which
highlight
lesion
region
information.
Thirdly,
(GIFM)
feature
flow
strategy
designed.
It
converts
problem
into
weight
assignment
extraction
realized
gradually.
Finally,
new
dataset
PET-CT
established,
contains
2575
pairs
images
(PCLset).
experimental
results
on
PCLset
show
compared
with
other
nine
models,
achieve
better
performance.
In
CT
window
PET
comparison
experiment,
SD,
IE,
AG,
QAB/F,
VIF
EI
indexes
improved
20.38
%,7.70
%,16.44
%,
21.90
%,11.52
%
33.95
respectively.
