IEEE Transactions on Pattern Analysis and Machine Intelligence,
Journal Year:
2023,
Volume and Issue:
45(9), P. 11040 - 11052
Published: April 19, 2023
Deep
learning
based
fusion
methods
have
been
achieving
promising
performance
in
image
tasks.
This
is
attributed
to
the
network
architecture
that
plays
a
very
important
role
process.
However,
general,
it
hard
specify
good
architecture,
and
consequently,
design
of
networks
still
black
art,
rather
than
science.
To
address
this
problem,
we
formulate
task
mathematically,
establish
connection
between
its
optimal
solution
can
implement
it.
approach
leads
novel
method
proposed
paper
constructing
lightweight
network.
It
avoids
time-consuming
empirical
by
trial-and-test
strategy.
In
particular
adopt
learnable
representation
task,
which
construction
guided
optimisation
algorithm
producing
model.
The
low-rank
(LRR)
objective
foundation
our
matrix
multiplications,
are
at
heart
transformed
into
convolutional
operations,
iterative
process
replaced
special
feed-forward
Based
on
an
end-to-end
constructed
fuse
infrared
visible
light
images.
Its
successful
training
facilitated
detail-to-semantic
information
loss
function
preserve
details
enhance
salient
features
source
Our
experiments
show
exhibits
better
state-of-the-art
public
datasets.
Interestingly,
requires
fewer
parameters
other
existing
methods.
IEEE Transactions on Image Processing,
Journal Year:
2020,
Volume and Issue:
29, P. 4980 - 4995
Published: Jan. 1, 2020
In
this
paper,
we
proposed
a
new
end-to-end
model,
termed
as
dual-discriminator
conditional
generative
adversarial
network
(DDcGAN),
for
fusing
infrared
and
visible
images
of
different
resolutions.
Our
method
establishes
an
game
between
generator
two
discriminators.
The
aims
to
generate
real-like
fused
image
based
on
specifically
designed
content
loss
fool
the
discriminators,
while
discriminators
aim
distinguish
structure
differences
source
images,
respectively,
in
addition
loss.
Consequently,
is
forced
simultaneously
keep
thermal
radiation
texture
details
image.
Moreover,
fuse
resolutions,
e.g.,
low-resolution
high-resolution
image,
our
DDcGAN
constrains
downsampled
have
similar
property
with
This
can
avoid
causing
information
blurring
or
detail
loss,
which
typically
happens
traditional
methods.
addition,
also
apply
multi-modality
medical
positron
emission
tomography
magnetic
resonance
qualitative
quantitative
experiments
publicly
available
datasets
demonstrate
superiority
over
state-of-the-art,
terms
both
visual
effect
metrics.
code
at
https://github.com/jiayi-ma/DDcGAN.
International Journal of Computer Vision,
Journal Year:
2020,
Volume and Issue:
129(1), P. 23 - 79
Published: Aug. 4, 2020
Abstract
As
a
fundamental
and
critical
task
in
various
visual
applications,
image
matching
can
identify
then
correspond
the
same
or
similar
structure/content
from
two
more
images.
Over
past
decades,
growing
amount
diversity
of
methods
have
been
proposed
for
matching,
particularly
with
development
deep
learning
techniques
over
recent
years.
However,
it
may
leave
several
open
questions
about
which
method
would
be
suitable
choice
specific
applications
respect
to
different
scenarios
requirements
how
design
better
superior
performance
accuracy,
robustness
efficiency.
This
encourages
us
conduct
comprehensive
systematic
review
analysis
those
classical
latest
techniques.
Following
feature-based
pipeline,
we
first
introduce
feature
detection,
description,
handcrafted
trainable
ones
provide
an
these
theory
practice.
Secondly,
briefly
typical
matching-based
understanding
significance
matching.
In
addition,
also
objective
comparison
through
extensive
experiments
on
representative
datasets.
Finally,
conclude
current
status
technologies
deliver
insightful
discussions
prospects
future
works.
survey
serve
as
reference
(but
not
limited
to)
researchers
engineers
related
fields.
IEEE Transactions on Image Processing,
Journal Year:
2020,
Volume and Issue:
29, P. 4733 - 4746
Published: Jan. 1, 2020
Image
decomposition
is
crucial
for
many
image
processing
tasks,
as
it
allows
to
extract
salient
features
from
source
images.
A
good
method
could
lead
a
better
performance,
especially
in
fusion
tasks.
We
propose
multi-level
based
on
latent
low-rank
representation(LatLRR),
which
called
MDLatLRR.
This
applicable
fields.
In
this
paper,
we
focus
the
task.
develop
novel
framework
MDLatLRR,
used
decompose
images
into
detail
parts(salient
features)
and
base
parts.
nuclear-norm
strategy
fuse
parts,
parts
are
fused
by
an
averaging
strategy.
Compared
with
other
state-of-the-art
methods,
proposed
algorithm
exhibits
performance
both
subjective
objective
evaluation.
IEEE Transactions on Instrumentation and Measurement,
Journal Year:
2020,
Volume and Issue:
70, P. 1 - 14
Published: Dec. 1, 2020
Visible
images
contain
rich
texture
information,
whereas
infrared
have
significant
contrast.
It
is
advantageous
to
combine
these
two
kinds
of
information
into
a
single
image
so
that
it
not
only
has
good
contrast
but
also
contains
details.
In
general,
previous
fusion
methods
cannot
achieve
this
goal
well,
where
the
fused
results
are
inclined
either
visible
or
an
image.
To
address
challenge,
new
framework
called
generative
adversarial
network
with
multiclassification
constraints
(GANMcC)
proposed,
which
transforms
multidistribution
simultaneous
estimation
problem
fuse
and
in
more
reasonable
way.
We
adopt
estimate
distributions
light
domains
at
same
time,
game
discrimination
will
make
result
balanced
manner,
as
addition,
we
design
specific
content
loss
constrain
generator,
introduces
idea
main
auxiliary
extraction
gradient
intensity
enable
generator
extract
sufficient
from
source
complementary
manner.
Extensive
experiments
demonstrate
advantages
our
GANMcC
over
state-of-the-art
terms
both
qualitative
effect
quantitative
metric.
Moreover,
method
can
even
overexposed.
Our
code
publicly
available
https://github.com/jiayi-ma/GANMcC.
