Applied Intelligence

https://doi.org/10.1007/s10489-022-03591-4

NOSMFuse: An infrared and visible image fusion approach based

on norm optimization and slime mold architecture
Shuai Hao1 · Tian He1 · Xu Ma1 · Beiyi An1 · Hu Wen2 · Feng Wang3

Accepted: 5 April 2022

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract
In existing infrared and visible image fusion algorithms, it is usually difficult to maintain a good balance of meaningful
information between two source images, which easily leads to the omission of important fractional information in a particular
source image. To address this issue, a novel fusion algorithm based on norm optimization and slime mold architecture,
called NOSMFuse, is proposed. First, an interactive information decomposition method based on mutually guided image
filtering is devised and utilized to obtain the corresponding base and detail layers. Subsequently, the differentiation feature
extraction operator is formulated and employed to fuse the base layers. In addition, we design a norm optimization-based
fusion strategy for the detail layers and a loss function that considers both the intensity fidelity and the gradient constraint.
Finally, to further balance the useful information of the base and detail layers contained in the fusion image, we propose
a slime mold architecture based image reconstruction method that generates fusion results through adaptive optimization.
The experimental results show that the proposed NOSMFuse is superior to 12 other state-of-art fusion algorithms, both
qualitatively and quantitatively.

Keywords Image fusion · Infrared and visible images · Norm optimization · Slime mold architecture

1 Introduction into two categories: traditional fusion algorithms and deep

learning-based fusion algorithms [6].
Infrared and visible image fusion technology is aimed In general, a traditional fusion algorithm measures
at generating a fusion image with prominent targets and the activity level of pixels or regions in the spatial or
rich textures, according to the formulated fusion strategy. transform domain, and finally realizes the image fusion
Currently, such image fusion tasks are widely utilized task according to a certain fusion rule. It mainly includes
in satellite remote sensing [1], image segmentation [2], multiscale-decomposition [7–9], sparse-representation [10–
medical image analysis [3], electrical fault diagnosis [4] and 12], subspace [13–15], saliency [16–18] and hybrid
other fields. algorithms [19–21]. Although the above fusion algorithms
Owing to their unique thermal radiation imaging, infrared have their own advantages, they must manually design
images generally exhibit a prominent contrast and insufficient appropriate decomposition or fusion rules, which leads to a
details. Conversely, visible images are generated according to weak generalization ability.
the principle of optical reflection, containing rich details and In recent years, many deep-learning-based algorithms
conforming to the human visual perception. However, they have emerged, including autoencoder- and generative
are easily affected by environmental factors, resulting in a adversarial network (GAN) based algorithms. Autoencoder-
relatively low image contrast [5]. To make the most of the based algorithms divide an image fusion task into an
complementary information between two source images, encoder, fusion layer, and decoder. The fusion layer is
infrared and visible image fusion tasks can be divided discarded during the training process. An encoder network
is utilized to extract the deep features, and the input images
are adaptively reconstructed using the decoder network.
 Xu Ma Based on this, we can determine the weights of the encoder
[email protected]
and decoder. During the test process, an appropriate fusion
strategy is then designed and employed for specific fusion
Extended author information available on the last page of the article. tasks to fuse the deep features. For example, Prabhakar
 S. Hao et al.

et al. built the Deepfuse model [22]. However, it is The main contributions of this study are summarized as
worth noting that the original networks cannot satisfy follows.
the corresponding extraction, fusion, and reconstruction
1. We propose an interactive information decomposition
requirements. Therefore, Li et al. successively presented the
method based on mutually guided image filtering
Densefuse [23], NestFuse [24], and RFN-Nest [25] fusion
(I2MuGIF). With this method, one source image
algorithms. The authors introduced an improved encoder
is regarded as a guidance image to decompose
with a dense block structure, an improved decoder with a
another source image. Based on the structure transfer
nest connection structure, and an improved fusion layer with
characteristics, the base and detail layer images
a residual fusion network to enhance the relatively simple
obtained are beneficial to subsequent fusion strategies.
network architecture, which can enhance the fusion effects
2. We formulated a differentiation feature extraction
of the aforementioned algorithms.
operator to fuse the base layers. The fusion strategy can
Inspired by the idea of the two-person game, GAN-
identify prominent perceptual visual structures, regions,
based algorithms can be divided into two parts: a generator
or objects by comparing the pixel differences between
and a discriminator. Here, a generator is used to generate
two source images and facilitate the reconstruction of
a fusion image that attempts to deceive the discriminator.
the global structure information in the base layers.
Meanwhile, a discriminator is trained to distinguish the
3. We describe the design of a norm-optimization-based
correlation between the generated fusion image and the
fusion strategy for the detail layers. With this strategy,
ground truth. During the training process, the corresponding
a loss function that considers both the intensity fidelity
generation and discrimination abilities can be enhanced
and gradient constraint is established. Subsequently, the
through an adversarial relationship. Typical algorithms
splitting Bregman algorithm is adopted to solve this
such as FusionGAN [26], DDcGAN [27], and PMGI [28]
optimization problem.
have an unsupervised distribution estimation ability suitable
4. To balance the meaningful information contained in
for image fusion tasks. However, these deep-learning-
the base and detail layers, we constructed an adap-
based algorithms still have the following weaknesses: 1)
tive objective function and introduced a slime mold
The number of input images directly affects the training
architecture to obtain the optimization parameters. The
accuracy and fusion performance, 2) owing to a lack of
experimental results show that our NOSMFuse algo-
ground truth, scholars must formulate loss functions to
rithm can simultaneously retain significant brightness
complete the fusion tasks, and 3) most existing algorithms
information and texture details.
ignore the balance of complementary information between
two source images, resulting in a lack of valuable The remainder of this paper is organized as follows.
information of a particular source image in the fusion results In Section 2, mutually guided image filtering is briefly
[29, 30]. reviewed. In Section 3, the proposed NOSMFuse algorithm
To overcome the above drawbacks, a novel fusion is described in detail. The experiments and analyses are
algorithm based on norm optimization and a slime mold presented in Sections 4, and 5 provides some concluding
architecture (NOSMFuse) is proposed. Figure 1 shows the remarks.
comparison results between the different fusion methods
and our NOSMFuse algorithm. The enlarged view of
the red rectangular boxes indicates that the proposed 2 Mutually guided image filtering
NOSMFuse can enhance the foreground target regions in the
infrared image while capturing more detailed background Mutually guided image filters (MuGIFs) [31] are robust
information from the visible image. joint image filters that can adaptively perceive the edges

Fig. 1 Comparison of the results between different fusion methods and NOSMFuse. The red boxes indicate that the comparison algorithms
our NOSMFuse algorithm. From left to right: infrared image; visible lose the intensity information and detailed features at different degrees,
image; fusion image of a traditional-based method, HMSD [9]; fusion whereas the proposed NOSMFuse better highlights prominent targets
image of an autoencoder-based method, NestFuse [24]; fusion image and has sufficient textures
of a GAN-based method, PMGI [28]; and fusion image of the proposed
NOSMFuse: An infrared and visible image fusion approach

of the input images. In this study, because it can is located in the marginal region, the penalty factor of
maintain a mutual structure and avoid misleading structural ∇d T (x, y) is small, and the gradient of the target image is
inconsistencies, MuGIF is employed to decompose the preserved. The following filtering rules can be formulated
infrared and visible images. In this section, the MuGIF based on the concept of relevant structures:
principle is introduced.
First, three types of image structures are defined: arg min σo R(T , I )+δo T −T0 22 +σi R(I, T )+δi I −I0 22
T ,I
mutual, inconsistent, and flat. At a particular pixel, the
mutual structure presents the pixel gradients of the two (2)
images, which are both sufficiently strong. The inconsistent where σo , δo , σi , and δi denote the nonnegative constants for
structure indicates that the pixel gradient of one image is balancing each term, where σo = σi = σ and δo = δi = δ,
strong and that of the other is weak. The flat structure and T − T0 22 and I − I0 22 denote the fidelity terms,
indicates that the pixel gradients of the two images are aiming to prevent trivial solutions by constraining T and I ,
both weak. Subsequently, a novel measurement, relevant respectively.
structure, is designed to evaluate the structural similarity Through a global optimization of (2), a filtered image can
between two input images, and a global optimization be obtained. For convenience, the mutually guided image-
problem is proposed to achieve the image filtering. filtering algorithm is denoted as muGI F (T , I, σ, δ) in this
The target image and guidance image are defined as study.
T (x, y) and I (x, y), respectively, the sizes of which are
m × n. The relevant structure can then be expressed as
follows: 3 Proposed NOSMFuse
  |∇d T (x, y)|
R(T , I ) = (1)
|∇d I (x, y)| A flowchart of the proposed NOSMFuse is shown in Fig. 2.
(x,y) d∈{h,v}
In this section, we describe the design of the proposed
where ∇d represents the first-order derivative; and h and v fusion algorithm in detail.
denote the horizontal and vertical directions, respectively. Step 1: I2MuGIF-based image decomposition
The relevant structure measures the structural differences Mutually guided image filtering has a good structural
between the target and the guidance images. If I (x, y) transfer characteristic, which can effectively smooth the

Fig. 2 Flowchart of NOSMFuse

 S. Hao et al.

inconsistent structure between two images. Therefore, generated and expressed as follows:
from the perspective of information complementarity 
between the two source images, an I2MuGIF-based W1 (x, y) = B1 (x, y) − B2 (x, y)
(5)
image decomposition method was designed. First, the W2 (x, y) = B2 (x, y) − B1 (x, y)
corresponding base layers of the infrared and visible images To reasonably allocate the weights, an adaptive fusion
were obtained using (3): strategy is proposed, as shown in (6).

B1 = muGI F (I R, V I S, σ, δ)
B2 = muGI F (V I S, I R, σ, δ)
(3) FB (x, y) = Wb (x, y)B1 (x, y) + (1 − Wb (x, y)) B2 (x, y)
(6)
where σ and δ denote the balance parameters, σ = 0.06,
δ = 1, and B1 and B2 are the filtered results of two source where
images, representing the base layers of the infrared image W1 (x, y) − W2 (x, y)
I R and the visible image V I S, respectively. Wb (x, y) = 0.5 + (7)
2
The corresponding detail layers can then be obtained
by (4). Here, if the weight values of pixels in B1 and B2 satisfy
 W1 = W2 , then Wb = 0.5 indicates that the fusion image
D1 = I R − B 1 of the base layer is equivalent to that obtained by weight
(4)
D2 = V I S − B 2 averaging. If W1 < W2 is satisfied, then Wb < 0.5 indicates
that the information contained in the base layer B2 will be
The base and detail layers are shown in Fig. 3. The
retained to a large extent. Otherwise, if Wb > 0.5 is used,
illustrated image details in the red boxes indicate that,
more information from the base layer B1 will be effectively
compared with the MuGIF-based image decomposition
preserved.
method (where the target image is consistent with the
Step 3: Fusion strategy of detail layers
guidance image), the proposed image decomposition
In Fig. 3, although the base and detail layers obtained can
method has a stronger expression ability for structure and
better express the global structure and detail information,
texture information.
respectively, the partial brightness information of the person
Step 2: Fusion strategy of base layers
is retained in the infrared detail layer, as shown in the yellow
The average-weight-based fusion strategy [32, 33] easily
box. To express the texture information of the two detail
leads to the loss of important contour information during
layers and compensate for the partial brightness information
the fusion process. In this study, to identify prominent
from the infrared detail layer, a fusion strategy of the detail
perceptual visual structures, regions, or objects in the base
layer based on a norm optimization was designed as follows:
layers, we constructed a novel feature operator, called
a differentiation feature extraction operator, based upon 1. Construction of the loss function
which an adaptive fusion strategy for the base layer was
The loss function is dependent on the terms of the intensity
designed.
fidelity and gradient constraint, which can be described as
By calculating the pixel difference between the two base
follows:
layers, the differentiation feature extraction operator can be
formulated and used to estimate the brightness regions in (1) The intensity fidelity term is aimed at maintaining the
each base image. The weight maps W1 and W2 are then light-dark ratio between the target and background in

Fig. 3 Images of the base and

detail layers. The four images in
the first row are the base and
detail layers obtained through
the mutually guided image
filtering without the interaction
of information. The four images
in the second row are the base
and detail layers obtained using
I2MuGIF-based image
decomposition
NOSMFuse: An infrared and visible image fusion approach

the infrared detail layer image, which can be defined To solve (12), we decouple it into two subproblems. In
as solving each sub-problem, the remaining variables are set
1 as fixed values. The subproblems are then iterated until
φ = FD − D1 22 (8) convergence is achieved. In addition, we initialize f 0 = z
2
(2) Because the image gradients tend to be sparse, the l1 and dx0 = dy0 = bx0 = by0 = 0.
norm is employed to formulate the gradient constraint, (1) Solution to f k+1
which has a better optimization characteristic. There- 
1 k ε 2
fore, the gradient constraint term can be defined as f k+1 = arg min f − z22 + d x − ∇ x f k − bx
f 2 2 2
 (14)
ε 2
ϕ = ∇FD − ∇D1 − ∇D2 11 (9) + d y − ∇ y f k − by
2 2
where ∇FD , ∇D1 , and ∇D2 denote the gradients of
In this study, the Gauss–Seidel iterative algorithm
the infrared detail layer, the visible detail layer, and the
is employed to solve the above equation.
[35]
fusion image of the detail layer, respectively.
(2) Solutions to (dxk+1 , dyk+1 ) and (bxk+1 , byk+1 )
Finally, the fusion image of the detail layer FD
can be constituted by constructing the following loss  ε
dxk+1 , dyk+1 = arg min μ (dx , dy ) 2
+ dx − ∇x f − bx 22
function: dx ,dy 2
  ε 2
1 + dy − ∇y f − by
arg min FD − D1 2 + μ ∇FD − ∇D1 − ∇D2 1
2 1 2 2

FD 2 (15)
(10) Note that dx and dy cannot be decoupled. We
where μ controls the trade-off between the intensity introduce a generalized shrinkage formula to explicitly
fidelity and gradient constraint. solve the minimization problem, which can be
expressed as follows:
2. Transformation of the loss function
μ ∇x f k + bxk
Letting FD = f + D1 + D2 and z = −D2 , then the (10) dxk+1 = max s k − ,0
ε sk
(16)
can then be expressed as follows: μ ∇y f k + byk
  dyk+1 = max s k − , 0
1 ε sk
arg min f − z2 + μ∇f 1
2 1
f 2 where
 (11) 
   2 sk = |∇x f k + bxk |2 + |∇y f k + byk |2 (17)
∇f 1 =
1
|∇fj | = (∇x f )2j + ∇y f j
j j Subsequently, the values of bxk+1 and byk+1 can be
where f is an intermediate image, j denotes the spatial obtained.
⎧
position of the image, ∇f represents the gradient of the ⎨ bk+1 = bk + ∇x f k+1 − d k+1 
x x x
intermediate image, and ∇x f and ∇y f denote the gradients ⎩ byk+1 = byk + ∇y f k+1 − dyk+1
(18)
of the horizontal and vertical directions, respectively.
Because the split Bregman method [34] can effectively deal 4. Reconstructing fusion image of the detail layer
with L1-regularized problems, it is adopted to optimize (11). By iterating the above subproblems until conver-
gence, the optimal result of the intermediate image f ∗ is
3. Solution to the minimization problem based on the split
obtained. The fusion images of the detail layer can then
Bregman algorithm
be reconstructed using a linear superposition, which can
Letting dx ≈ ∇x f and dy ≈ ∇y f , the split Bregman be expressed as follows:
formula of (11) can be expressed as follows:
F D = f ∗ + D1 + D2 (19)
 
1
f k+1 , dxk+1 , dyk+1 = arg min f − z22 + μ (dx , dy ) 2 Step 4: Slime mold architecture based image reconstruction
f,dx ,dy 2
(12) To further balance the meaningful information contained
ε ε 2
+ dx − ∇x f − bx 22 + dy − ∇y f − by 2 in the fusion image of the base and detail layers, a slime
2 2
mold architecture based image reconstruction algorithm is
ε 2 ε
where dx − ∇x f − bx 22 and dy − ∇y f − by 2 denote the proposed. After multiple iterations, the final fusion result
2 2
added penalty items, and dx , dy 2 can be expressed as (13). F can be generated through an adaptive optimization, as
 shown in the following formula:
    2
dx , dy 2
= (dx )2j + dy j (13)
j F = θ · FB + FD (20)
 S. Hao et al.

where θ denotes the optimization parameter. The Because the slime mold architecture can effectively solve
optimization design process can be described as follows: complex global optimization problems [37], we introduce
it to optimize parameter θ . The slime mold architecture
1. Construction of objective function
mainly explores the diffusion and foraging process of
The objective function consists of two parts, Lssim and Lcon , slime mold, which approaches food according to the odor
which can be expressed as follows: concentration. The higher the concentration is, the faster
the approaching speed. Moreover, in this architecture, two
min {λ · Lssim + Lcon } (21)
random oscillation parameters are introduced, which range
where Lssim and Lcon are the structural similarity and content from [−a, a] and [−1, 1], respectively. With an increase in
losses between the fused and source images, respectively; the number of iterations, their values eventually decrease to
and λ denotes a balance parameter. In this study, λ is set to zero. The value of a is given by (26). The selective behavior
1e0, 1e1, 1e2, and 1e3, respectively. of the slime mold can be simulated through the synergistic
effect between these two parameters.
(1) Structural similarity loss function:    
t
Lssim = 2 − ssim(F, I R) − ssim(F, V I S) (22) a = arctan h − +1 (26)
max− t
where ssim(•) denotes the structural similarity index where t is the current iteration, and max− t denotes the
[36]. The larger the value, the more structural maximum number of iterations.
information from the source image is retained in the To update the position of the slime mold, (27) can be
fused image. constructed as follows:
(2) Content loss function ⎧
⎪
⎪ rand · (U B − LB) + LB, rand < z
The content loss function follows the master– −− → ⎨ −−→ − → − → −−→ −−→
θ new
= θb (t) + vb · W · θA (t) − θB (t) , r < p (27)
slave information and can be divided into two ⎪
⎪ −→
⎩−→
vc · θ (t), r ≥ p
parts: intensity information and gradient information.
Intensity information mainly exists in an infrared where LB and U B are the upper and lower bounds of the
image and is partially retained in a visible image. search scope, respectively; rand and r denote a random
The intensity loss function can then be formulated as −
→
number in [0, 1]; θ represents the current position of the
follows: −
→
slime mold; θb denotes the individual location with the
−
→ −
→
Lint = Lmain highest odor concentration; θA and θB are the two randomly
int + Lint = β0 · F − I R2 + β1 · F − V I S2 (23)
aux 2 2
selected positions from the slime mold; and p represents the
where Lmain
int and Laux
int represent the main and auxiliary segment point, the value of which can be expressed as (28).
intensity loss terms, respectively; and β0 and β1 denote −
→
In addition, W denotes the weight of the slime mold, which
2
the balance parameters, where in this study, β0 = and can be expressed as (29).
3
1
β1 = . p = tanh |S(i) − DF | (28)
3
Gradient information mainly exists within a visible −
→
where S(i) is the fitness of and i ∈ 1, 2, . . . , n; in addition,
θ,
image and partially remains in an infrared image, which can DF denotes the best fitness obtained in all iterations.
be represented through the following gradient loss function: ⎧  
⎪
⎪ bF − S(i)
⎨ 1 + r · log + 1 , condition
Lgrad = grad + Lgrad
Lmain aux
= β2 · ∇F − ∇V I S11 + β3 · ∇F − ∇I R11 (24) −−−−−−−−−−−−−→  bF − wF 
W (SmellI ndex(i)) = bF − S(i)
⎪
⎪
⎩ 1 − r · log + 1 , others
where Lmain aux
grad and Lgrad represent the main and auxiliary bF − wF
gradient loss terms, respectively; and β2 and β3 denote the (29)
2 1
balance parameters, where β2 = and β3 = .
3 3
In summary, the content loss function is composed of the
SmellI ndex = sort (S) (30)
following four parts:
Lcon = α · β0 · F − I R22 + β1 · F − V I S22
where condition indicates that S(i) is ranked within the first
(25) half of the population; bF and wF represent the best and
+ β2 · ∇F − ∇V I S11 + β3 · ∇F − ∇I R11 worst fitness levels obtained during the current iteration
where α denotes the parameter controlling the trade-off process, respectively; and SmellI ndex denotes the sorted
between the intensity and gradient loss, and in this study, sequence of fitness values.
α=
1
. After multiple iterations, the objective function can be
500 solved effectively and the fusion image F can be obtained,
2. Adaptive optimization based on slime mold architecture as shown in Fig. 4.
NOSMFuse: An infrared and visible image fusion approach

Fig. 4 Processed images

fusion image of fusion image of

fusion result
base layers detail layers

4 Experiments and analysis texture, to varying degrees. FusionGAN and RFN-Nest

are prone to generating fuzzy fusion results. Moreover,
To verify the superiority of the proposed NOSMFuse the fusion images obtained by the other three comparison
algorithm, 21 pairs of source images are randomly algorithms are biased toward visible images, which contain
selected from the TNO dataset1 and then qualitatively and rich textures but lose the thermal radiation information.
quantitatively analyzed. Moreover, an ablation experiment Compared with the 12 fusion methods above, the proposed
is introduced in detail. algorithm achieves a good balance between preserving the
details and highlighting the targets.
4.1 Qualitative analysis Figure 6 shows the qualitative fusion results for the
image sequence “Nato” in the TNO dataset when using
We selected the following 12 state-of-art fusion algorithms the different methods. The enlarged views of the red
for the comparative experiments: anisotropic diffusion- rectangular boxes indicate that, compared with the other
based fusion (ADF) [38], fourth order partial differential 12 classical fusion algorithms, the proposed algorithm has
equations (FPDE) [39], infrared feature extraction and visual the following advantages. First, NOSMFuse maintains the
information preservation (IFE-VIP) [40], hybrid multi-scale major thermal radiation information in the original infrared
decomposition with Gaussian and bilateral filters (HMSD) image, which can effectively distinguish the target from
[9], deep learning framework-based method (DeepFuse) the background. Second, our algorithm effectively enhances
[22], generate adversarial network fusion (FusionGAN) [26], the edge contour information of the pedestrian. In addition,
convolutional neural network-based method (IFCNN) [41], the proposed algorithm also improves the image definition,
nest connection-based fusion (NestFuse) [24], proportional making it more suitable for human perception systems.
maintenance of gradient and intensity (PMGI) [28], residual As shown in Fig. 7, the qualitative experiment results
fusion network-based method (RFN-Nest)[25], symmetric illustrate that the IFE-VIP, HMSD, IFCNN, NestFuse,
encoder-decoder with residual block (SEDRFuse) [42], and SEDRFuse, and STDFusionNet fusion models have better
salient target detection-based method (STDFusionNet) [43]. fusion effects than the other six comparison algorithms,
Partial comparison results are shown in Figs. 5-8. To facilitate with clearer fused images and more significant light-dark
the observation and subsequent analysis, the differences ratios. However, as shown in the red boxes in Fig. 7,
between the obtained fusion images are highlighted visually the details of the shrubs are partially lost owing to
in the red boxes. the introduction of too much infrared image interference
In Fig. 5, the fusion images obtained using ADF and information. By contrast, the proposed NOSMFuse not only
FPDE can express the global structure information well; has excellent image clarity, it also highlights various details.
however, the tree texture features cannot be effectively Figure 8 illustrates the qualitative fusion results on the
preserved from a visible image. The IFE-VIP and HMSD image sequence “Bunker.” The results of the comparative
fusion methods focus more on the energy information analysis show that the fusion images obtained by ADF,
extracted from the infrared image, resulting in the halo FPDE, DeepFuse, and RFN-Nest can express global
phenomenon around the contour of the tree and person, structural information; however, the images are generally
respectively. DeepFuse, IFCNN, and PMGI retain rich dark, which is not conducive to an observation. IFE-VIP,
information during the fusion process but still lose the FusionGAN, and STDFusionNet pay too much attention to
local features, such as chimney brightness and tree the brightness features from the original infrared images
while ignoring the equally important texture information
1 Available at: https://figshare.com/articles/dataset/TNO Image in the visible images. In addition, although HMSD,
Fusion Dataset/1008029 IFCNN, NestFuse, PMGI, and SEDRFuse can generate
 S. Hao et al.

Fig. 5 Experiment on “Kaptein” images. The last four images are the fusion results obtained using the proposed method with different SSIM
weights

Fig. 6 Experiment on “Nato” images. The last four images are the fusion results obtained using the proposed method with different SSIM weights

Fig. 7 Experiment on “Lake” images. The last four images are the fusion results obtained using the proposed method with different SSIM weights
NOSMFuse: An infrared and visible image fusion approach

Fig. 8 Experiment on “Bunker” images. The last four images are the fusion results obtained using the proposed method with different SSIM
weights

high-definition fused images, some local details such as local variance (QILV) [45], and spatial–spectral entropy-
the eave outlines and tree textures cannot be effectively based quality (SSEQ) [46] as quantitative evaluation indexes.
preserved. Comparatively, our NOSMFuse achieved a better For a fusion image, AG reflects its relative definition. SF
fusion performance in terms of the detail extraction, energy emphasizes its gradient distribution in the spatial domain.
preservation, and perceptual quality. The MSE measures the difference between the fusion and
reference images. The PSNR embodies this distortion by
4.2 Quantitative analysis comparing the peak and noise powers. The SSEQ utilizes its
local space and spectral entropy properties to evaluate the
To objectively evaluate the advantages of the proposed image quality across multiple distortion categories. QILV
algorithm, we selected average gradient (AG) [44], space assesses its non-stationarity from the perspective of the
frequency (SF) [44], mean squared error (MSE) [44], peak image structure. In addition, the better the fusion image is,
signal to noise ratio (PSNR) [44], quality index based on the larger the values of AG, SF, PSNR, and QILV, and the

Table 1 Average values of evaluation indexes for 21 fusion images

Algorithm AG SF MSE PSNR SSEQ QILV

ADF 4.4095 10.6449 0.0299 63.7302 21.3974 0.4073

FPDE 4.5988 10.8633 0.0316 63.5894 20.6330 0.4119
IFE-VIP 3.9792 10.6717 0.0475 61.6119 20.0807 0.4911
HMSD 4.0813 10.5153 0.0337 63.2061 20.3415 0.5982
DeepFuse 3.5130 8.9276 0.0352 63.0782 19.6942 0.4318
FusionGAN 2.2052 5.7973 0.0589 61.0455 25.2369 0.4396
IFCNN 4.3362 11.5035 0.0313 63.6480 21.1234 0.4856
NestFuse 3.6609 9.8091 0.0365 62.9183 20.5225 0.4723
PMGI 3.5725 8.7195 0.0359 62.9313 25.5850 0.4522
RFN-Nest 2.7337 6.1341 0.0395 62.6282 35.4359 0.4287
SEDRFuse 3.9556 10.1095 0.0329 63.3118 18.5788 0.4766
STDFusionNet 4.1479 11.6361 0.0454 62.0189 19.6839 0.4504
Ours 4.7086 13.4283 0.0296 63.8088 16.8704 0.4898
4.7085 13.4281 0.0295 63.8108 16.8417 0.4899
4.7127 13.4391 0.0293 63.8586 16.9015 0.4919
4.7197 13.4560 0.0290 63.9011 16.8663 0.4936
 S. Hao et al.

Fig. 9 Qualitative comparison

of our algorithm with four
ablation methods

smaller the values of MSE and SSEQ. The average values that NOSMFuse achieves the best fusion effects compared
of these evaluation indexes for 21 fusion images are shown to the other 12 fusion methods.
in Table 1. Note that the best values are indicated in red, and
the second-best values are indicated in blue. 4.3 Ablation experience
As shown in Table 1, the proposed NOSMFuse maintains
the optimal values for AG, SF, MSE, PSNR, and SSEQ. The To illustrate the superiority of the constructed image
largest AG and SF indicate that, through NOSMFuse, the decomposition method, special fusion strategies for the
fusion image obtained has the best definition and strongest base and detail layers, and the proposed slime mold
ability to express details. Meanwhile, the best MSE, PSNR, architecture-based image reconstruction algorithm, we
and SSEQ indicate that the fusion image obtained by conducted four ablation experiments using qualitative and
NOSMFuse not only contains more information, it also has quantitative evaluations. The four ablation experiments
the least distortion. Moreover, when λ = 1e2, 1e3, the QILV include a MuGIF-based fusion model without information
of the proposed NOSMFuse has the second-highest value, interaction (Without-I2MuGIF), a fusion model without
indicating that the structure of the fusion image is relatively the differentiation feature extraction based base layer
stable. fusion strategy (Without-DFE), a fusion model without
In conclusion, the results of the quantitative experiment the norm optimization based detail layer fusion strategy
are consistent with the qualitative analysis, which indicates (Without-NO), and a fusion model without the slime
NOSMFuse: An infrared and visible image fusion approach

Fig. 10 Quantitative MSE PSNR

comparison of our algorithm 0.25 70
with four ablation methods on Without-I2MuGIF: 0.0308 Without-DFE: 0.0324
four evaluation indexes Without-NO: 0.0384 Without-SMA: 0.0754
0.20 65
Ours: 0.0290

0.15 60

0.10 55

Without-I2MuGIF: 63.5957 Without-DFE: 63.5117

0.05 50
Without-NO: 62.6940 Without-SMA: 59.7777
Ours: 63.9011
0.00 45
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

SSEQ QILV
70 1.00
Without-I2MuGIF: 15.6882 Without-I2MuGIF: 0.4379 Without-DFE: 0.4264
60 Without-DFE: 16.0840 Without-NO: 0.2902 Without-SMA: 0.4781
Without-NO: 20.5227 0.75 Ours: 0.4936
50
Without-SMA: 19.5224
40 Ours: 16.8663
0.50
30

20
0.25
10

0 0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

mold architecture-based image reconstruction (Without- and detail information in a fusion image while ensuring a
SMA). It is worth noting that each fusion model in the richness of the information.
ablation experiments differs from the proposed NOSMFuse
in only a single aspect. For example, the DFE, NO, 2. Quantitative analysis To objectively verify the necessity of
and SMA modules are retained in the Without-I2MuGIF the proposed algorithm, we selected four evaluation indices,
fusion model. To fully demonstrate the effectiveness of our i.e., MSE, PSNR, SSEQ, and QILV, to conduct qualitative
algorithm, we compared it with four ablation experiments analyses of the ablation experiments. The comparisons are
on 21 pairs of source images from the TNO dataset, four of presented in Fig. 10. The proposed NOSMFuse maintains
which are shown in Fig. 9. the optimal average values of MSE, PSNR, and QILV.
Moreover, although the SSEQ cannot obtain the best value,
1. Qualitative analysis According to the observations in Fig. 9, it achieves close to the best value, and the local optimal
the following conclusions can be drawn: First, the fusion value is expressed in some of the source images, such as
images obtained by Without-I2MuGIF have more noise groups 7 and 13. Therefore, the results of the qualitative
than the proposed NOSMFuse, which demonstrates the and quantitative analyses of the ablation experiments
effectiveness of the interactive information decomposition demonstrate that each part of our method is necessary.
method based on mutually guided image filtering. Second,
although the Without-DFE fusion method can maintain the
overall brightness of the fusion image, the light-dark ratio 5 Conclusion
between the local feature region and the background region
will be lost to a certain extent during the fusion process. In this study, a novel fusion algorithm, NOSMFuse, was
Third, the Without-NO fusion method cannot effectively proposed. The proposed algorithm can effectively solve
retain the rich texture information in the source image, the problem in which most existing algorithms find it
particularly in the last column image. Finally, the Without- difficult to maintain a good balance of inherent information
SMA fusion method focuses more on the brightness in the source images. To enhance the expression ability
information, which easily leads to a serious disequilibrium of the structure and texture information, an I2MuGIF-
between the intensity and detail information. Compared to based image decomposition method was presented, and the
these four ablation experiments, the proposed NOSMFuse differentiation feature extraction operator was formulated
can maintain an effective balance between the brightness to fuse two base layer images. From the perspective of
 S. Hao et al.

the intensity fidelity and gradient constraint, we then decomposition with gaussian and bilateral filters. Information
proposed a detailed layer fusion strategy based on norm Fusion 30:15–26. https://doi.org/10.1016/j.inffus.2015.11.003
10. Liu Y, Chen X, Ward RK, Wang ZJ (2016) Image fusion with
optimization and utilized the split Bregman method to
convolutional sparse representation. IEEE Signal Process Lett
optimize the constructed loss function. Finally, an adaptive 23(12):1882–1886. https://doi.org/10.1109/LSP.2016.2618776
objective function was designed to further balance the useful 11. Kim M, Han DK, Ko H (2016) Joint patch clustering-based
information contained in the fusion image of the base and dictionary learning for multimodal image fusion. Information
Fusion 27:198–214. https://doi.org/10.1016/j.inffus.2015.03.003
detail layers. In addition, a slime mold architecture was
12. Li H, He X, Yu Z, Luo J (2020) Noise-robust image fusion with
introduced to reconstruct a fusion image. low-rank sparse decomposition guided by external patch prior. Inf
Experimental results show that the proposed algorithm Sci 523:14–37. https://doi.org/10.1016/j.ins.2020.03.009
not only highlights the important targets, it also effectively 13. Milgrom B, Avrahamy R, David T, Caspi A, Golovachev Y,
Engelberg S (2020) Extended depth-of-field imaging employ-
retains the rich details in the two source images.
ing integrated binary phase pupil mask and principal compo-
Acknowledgements This work was supported by the National Natural nent analysis image fusion. Opt Express 28(16):23862–23873.
Science Foundation of China [grant number 51804250], China https://doi.org/10.1364/OE.393037
Postdoctoral Science Foundation [grant number 2019M653874XB, 14. Wang Z, Deller Jr JR, Fleet BD (2016) Pixel-level multisensor
2020M683522], Scientific Research Program of Shaanxi Provincial image fusion based on matrix completion and robust principal
Department of Education [grant number 18JK0512], Natural Science component analysis. Journal of Electronic Imaging 25(1):013007.
Basic Research Program of Shaanxi [grant number 2021JQ-572, https://doi.org/10.1117/1.JEI.25.1.013007
2020JQ-757], Innovation Capability Support Program of Shaanxi 15. Fu Z, Wang X, Xu J, Zhou N, Zhao Y (2016) Infrared and visible
[grant number 2020TD-021], Xi ’an Beilin District Science and images fusion based on rpca and nsct. Infrared Phys Technol
Technology Project [grant number GX2116] and Weinan Science and 77:114–123. https://doi.org/10.1016/j.infrared.2016.05.012
Technology Project [grant number 2020ZDYF-JCYJ-196]. 16. Bavirisetti DP, Dhuli R (2016) Two-scale image fusion of visible
and infrared images using saliency detection. Infrared Phys
Technol 76:52–64. https://doi.org/10.1016/j.infrared.2016.01.009
Declarations 17. Ma J, Zhou Z, Wang B, Zong H (2017) Infrared and visible
image fusion based on visual saliency map and weighted
Competing interests The authors declare that there is no potential least square optimization. Infrared Phys Technol 82:8–17.
conflict of interest. https://doi.org/10.1016/j.infrared.2017.02.005
18. Zhao J, Gao X, Chen Y, Feng H, Wang D (2016) Multi-
window visual saliency extraction for fusion of visible
and infrared images. Infrared Phys Technol 76:295–302.
References https://doi.org/10.1016/j.infrared.2016.01.020
19. Chen G, Li L, Jin W, Zhu J, Shi F (2019) Weighted sparse repre-
1. Azarang A, Kehtarnavaz N (2021) A generative model sentation multi-scale transform fusion algorithm for high dynamic
method for unsupervised multispectral image fusion in range imaging with a low-light dual-channel camera. Opt Express
remote sensing. Signal Image Video Processing, 1–9. 27(8):10564–10579. https://doi.org/10.1364/OE.27.010564
https://doi.org/10.1007/s11760-021-01950-1 20. Mao Q, Zhu Y, Lv C, Lu Y, Yan X, Wei D, Yan S, Liu
2. Ding Z, Wang T, Sun Q, Wang H (2021) Adaptive fusion with J (2020) Image fusion based on multiscale transform and
multi-scale features for interactive image segmentation. Appl sparse representation to enhance terahertz images. Opt Express
Intell, 1–12. https://doi.org/10.1007/s10489-020-02114-3 28(17):25293–25307. https://doi.org/10.1364/OE.396604
3. Du J, Li W, Tan H (2020) Three-layer medical image 21. Shibu DS, Priyadharsini SS (2021) Multi scale decomposition
fusion with tensor-based features. Inf Sci 525:93–108. based medical image fusion using convolutional neural network
https://doi.org/10.1016/j.ins.2020.03.051 and sparse representation. Biomedical Signal Processing and
4. Ge Z, Jiang X, Tong Z, Feng P, Zhou B, Xu M, Wang Z, Pang Control 69:102789. https://doi.org/10.1016/J.BSPC.2021.102789
Y (2021) Multi-label correlation guided feature fusion network 22. Ram Prabhakar K, Sai Srikar V, Venkatesh Babu R (2017)
for abnormal ecg diagnosis. Knowl-Based Syst 233:107508. Deepfuse: a deep unsupervised approach for exposure fusion
https://doi.org/10.1016/j.knosys.2021.107508 with extreme exposure image pairs. In: Proceedings of the IEEE
5. Hou R, Zhou D, Nie R, Liu D, Xiong L, Guo Y, Yu C (2020) international conference on computer vision, pp 4714–4722.
Vif-net: an unsupervised framework for infrared and visible image https://doi.org/10.1109/ICCV.2017.505
fusion. IIEEE Transactions on Computational Imaging 6:640–651. 23. Li H, Wu X-J (2018) Densefuse: a fusion approach to infrared
https://doi.org/10.1109/TCI.2020.2965304 and visible images. IEEE Trans Image Process 28(5):2614–2623.
6. Jiang L, Fan H, Li J (2021) A multi-focus image fusion method https://doi.org/10.1109/TIP.2018.2887342
based on attention mechanism and supervised learning. Appl 24. Li H, Wu X-J, Durrani T (2020) Nestfuse: an infrared and
Intell, 1–19. https://doi.org/10.1007/s10489-021-02358-7 visible image fusion architecture based on nest connection and
7. Hu Z, Liang W, Ding D, Wei G (2021) An improved spatial/channel attention models. IEEE Trans Instrum Meas
multi-focus image fusion algorithm based on multi- 69(12):9645–9656. https://doi.org/10.1109/TIM.2020.3005230
scale weighted focus measure. Appl Intell, 1–17. 25. Li H, Wu X-J, Kittler J (2021) Rfn-nest: an end-to-end residual
https://doi.org/10.1007/s10489-020-02066-8 fusion network for infrared and visible images. Information
8. Dinh P.-H. (2021) Multi-modal medical image fusion based on Fusion 73:72–86. https://doi.org/10.1016/j.inffus.2021.02.023
equilibrium optimizer algorithm and local energy functions. Appl 26. Ma J, Yu W, Liang P, Li C, Jiang J (2019) Fusion-
Intell, 1–16. https://doi.org/10.1007/s10489-021-02282-w gan: a generative adversarial network for infrared and
9. Zhou Z, Wang B, Li S, Dong M (2016) Perceptual fusion visible image fusion. Information Fusion 48:11–26.
of infrared and visible images through a hybrid multi-scale https://doi.org/10.1016/j.inffus.2018.09.004
NOSMFuse: An infrared and visible image fusion approach

27. Ma J, Xu H, Jiang J, Mei X, Zhang X-P (2020) Ddcgan: 43. Ma J, Tang L, Xu M, Zhang H, Xiao G (2021) Stdfusion-
a dual-discriminator conditional generative adversarial network net: an infrared and visible image fusion network based on
for multi-resolution image fusion. IEEE Trans Image Process salient target detection. IEEE Trans Instrum Meas 70:1–13.
29:4980–4995. https://doi.org/10.1109/TIP.2020.2977573 https://doi.org/10.1109/TIM.2021.3075747
28. Zhang H, Xu H, Xiao Y, Guo X, Ma J (2020) Rethinking 44. Ma J, Ma Y, Li C (2019) Infrared and visible image fusion
the image fusion: a fast unified image fusion network based methods and applications: a survey. Information Fusion 45:153–
on proportional maintenance of gradient and intensity. In: 178. https://doi.org/10.1016/j.inffus.2018.02.004
Proceedings of the AAAI conference on artificial intelligence, 45. Aja-Fernandez S, Estepar RSJ, Alberola-Lopez C, Westin
pp 12797–12804. https://doi.org/10.1609/aaai.v34i07.6975 C.-F. (2006) Image quality assessment based on local
29. Li G, Lin Y, Qu X (2021) An infrared and visible variance. In: International conference of the IEEE engi-
image fusion method based on multi-scale transformation neering in medicine and biology society, pp 4815–4818.
and norm optimization. Information Fusion 71:109–129. https://doi.org/10.1109/IEMBS.2006.259516
https://doi.org/10.1016/j.inffus.2021.02.008 46. Liu L, Liu B, Huang H, Bovik AC (2014) No-reference
30. Tan M-J, Gao S-B, Xu W-Z, Han S-C (2020) Visible-infrared image quality assessment based on spatial and spectral
image fusion based on early visual information processing mech- entropies. Signal Process Image Commun 29(8):856–863.
anisms. IEEE Trans Circuits Syst Video Technol 31(11):4357– https://doi.org/10.1016/j.image.2014.06.006
4369. https://doi.org/10.1109/TCSVT.2020.3047935
31. Guo X, Li Y, Ma J, Ling H (2018) Mutually guided image
filtering. IEEE Trans Pattern Anal Mach Intell 42(3):694–707. Publisher’s note Springer Nature remains neutral with regard to
https://doi.org/10.1109/TPAMI.2018.2883553 jurisdictional claims in published maps and institutional affiliations.
32. Li H, Wu X-J, Kittler J (2020) Mdlatlrr: a novel decomposition
method for infrared and visible image fusion. IEEE Trans Image
Process 29:4733–4746. https://doi.org/10.1109/TIP.2020.2975984
33. Bavirisetti DP, Xiao G, Zhao J, Dhuli R, Liu G (2019) Multi-
scale guided image and video fusion: a fast and efficient approach.
Circuits, Systems, and Signal Processing 38(12):5576–5605.
https://doi.org/10.1007/s00034-019-01131-z Shuai Hao was born in 1986.
34. Goldstein T, Osher S (2009) The split bregman method for He received his Ph.D. degree
l1-regularized problems. SIAM Journal on Imaging Sciences from School of Automation,
2(2):323–343. https://doi.org/10.1137/080725891 Northwestern Polytechnical
35. Bakari A, Dahiru I (2018) Comparison of jacobi and gauss- University in 2014. He is cur-
seidel iterative methods for the solution of systems of lin- rently an assistant professor
ear equations. Asian Research Journal of Mathematics, 1–7. in Xi’an University of Science
https://doi.org/10.9734/ARJOM/2018/34769 and Technology. So far he
36. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) has published more than 30
Image quality assessment: from error visibility to struc- papers. His research interest is
tural similarity. IEEE Trans Image Process 13(4):600–612. vision navigation and digital
https://doi.org/10.1109/TIP.2003.819861 image processing.
37. Li S, Chen H, Wang M, Heidari AA, Mirjalili S (2020)
Slime mould algorithm: a new method for stochastic
optimization. Futur Gener Comput Syst 111:300–323.
https://doi.org/10.1016/j.future.2020.03.055
38. Bavirisetti DP, Dhuli R (2015) Fusion of infrared and
visible sensor images based on anisotropic diffusion and
karhunen-loeve transform. IEEE Sensors J 16(1):203–209.
https://doi.org/10.1109/JSEN.2015.2478655
39. Bavirisetti DP, Xiao G, Liu G (2017) Multi-sensor image fusion
based on fourth order partial differential equations. In: 20Th
international conference on information fusion (fusion), pp 1–9. Tian He received the B.S.
https://doi.org/10.23919/ICIF.2017.8009719 degree in electrical engineer-
40. Zhang Y, Zhang L, Bai X, Zhang L (2017) Infrared and visual ing and its automation from
image fusion through infrared feature extraction and visual Anhui Polytechnic University,
information preservation. Infrared Phys Technol 83:227–237. Wuhu, China, in 2020. She
https://doi.org/10.1016/j.infrared.2017.05.007 is currently pursuing the M.S.
41. Zhang Y, Liu Y, Sun P, Yan H, Zhao X, Zhang L degree in electrical engineer-
(2020) Ifcnn: a general image fusion framework based on ing with the Xi’an University
convolutional neural network. Information Fusion 54:99–118. Of Science And Technology,
https://doi.org/10.1016/j.inffus.2019.07.011 Xi’an. Her current research
42. Jian L, Yang X, Liu Z, Jeon G, Gao M, Chisholm D (2020) interests include image fusion,
Sedrfuse: a symmetric encoder-decoder with residual block deep learning and image pro-
network for infrared and visible image fusion. IEEE Trans Instrum cessing.
Meas 70:1–15. https://doi.org/10.1109/TIM.2020.3022438
 S. Hao et al.

Xu Ma was born in 1985. She Hu Wen received the Ph.D.

received her Ph.D. degree degree from Xi’an University
from School of Automation, of Science And Technology,
Northwestern Polytechni- Xi’an, China, in 2003. He is
cal University in 2017. She currently the deputy dean of
is currently a lecturer in the college of energy and engi-
Xi’an University of Science neering in Xi’an University
and Technology. So far she of Science and Technology,
has published more than 15 Xi’an, China. He was selected
papers. Her research interest is as the “Hundred-Thousand-Ten
vision navigation and digital Thousand Project” in 2015.
image processing. His current research interests
include mine emergency res-
cue theory, emergency rescue
visualization system and coal
fire disaster prevention and
control.

Beiyi An was born in Rizhao, Feng Wang was born in1981.

Shandong, in 1997. She He received her Ph.D. degree
received the B.S. degree in from School of Automation,
electrical engineering and auto- Northwestern Polytechnical
mation from Xi’an University University in 2019. He is
of Science and Technology, currently a lecturer in Weinan
Xian, China, in 2020. She is Normal University. So far he
currently pursuing the M.S. has published more than 9
degree in electrical engineer- papers. His research interest is
ing and automation from remote sensing image change
Xi’an University of Science detection, deep learning,
and Technology. Her current image fusion.
research interests include
image processing, image
fusion and deep learning.

Affiliations
Shuai Hao1 · Tian He1 · Xu Ma1 · Beiyi An1 · Hu Wen2 · Feng Wang3

Shuai Hao
[email protected]
Tian He
[email protected]
Beiyi An
[email protected]
Hu Wen
[email protected]
Feng Wang
[email protected]
1 College of Electrical and Control Engineering, Xi’an University
of Science and Technology, Xi’an, 710054, China
2 College of Safety and Engineering, Xi’an University of Science
and Technology, Xi’an, 710054, China
3 School of Physics and Electrical Engineering, Weinan Normal
University, Weinan, 71400, China