[Link] / [Link].

6(4) (2024) ISSN: 2663-2187

[Link] 10.33472/AFJBS.6.Si3.2024.1115-1126

Hybrid Fusion Based Target Detection Methods Visible and

Images Using Guided Filters
1
[Link], 2Santhalakshmi M, 3Shanthakumar M, 4Janarthanam S
123
Assistant Professors, Department of Computer Science, Kamban College of Arts
and Science, Coimbatore.
4
Associate Professor, School of Science and Computer Studies, CMR University,
Bengaluru.
Corresponding Author : professorshanth@[Link]

Abstract
The combination of visible and infrared (IR) images to
create an absolute, accurate, and dependable image is the
most widely used technique in image processing
applications. This paper proposes a method called Two-
Scale Decomposition (TSD) with Guided Filtering (GF)
Volume:6,IssueSi3,2024 and Phase Congruency (PC) and Sum Modified Laplacian
(SML) has been developed. TSD is used in TSD-PS-GF
Received:14April2024 to decompose the IR and VIS pictures, respectively, and
obtain the base and detail layers to use TSD-PS with
Accepted: 11May2024
Sturdy GF (TSD-PS-SGF) rather than GF to improve the
image fusion's resilience. The Enhanced Preconditioned
Conjugate Gradient (EPCG) method is used for each SGF
doi: iteration to optimize the conjugate routes and RLS. This
method handles irregularities in the structure and achieves
10.33472/AFJBS.6.Si3.2024.1115-
a high rate of convergence.
1126
Keywords : Infrared, Visible, Two-Scale Decomposition,
TSD-PS-SGF, TSD-PS-ASGF, TSD-DNN-ASGF
 [Link] / [Link]. 6(4) (2024) Page 1116 of 12

Introduction
Target identification is a very crucial technology for military operations that has not
yet realized its complete tactical secure. It is the process of recognizing the potential
military target as being a specific target. Completely reliable automated target
identification can improve the lethality and survivability of the war fighter and
platform. Automated target identification performs on sensor information for
processing data to make the decision. The primary value added to a weapons system
of automated target identification is engagement timeline reduction for targets
acquisition. The rapid acquisition and servicing of targets increase lethality and
survivability of the weapons platform.
Persistent Surveillance presents an earlier military application chance for providing
lower technically sophisticated automated target identification. Automated target
identification can provide significant enhancement to military weapons platforms over
human-only performance. Also, can provide enhancements to the weapons operator or
intelligence analyst for fire control, surveillance, reconnaissance, intelligence, PS and
situational awareness provide completely autonomous target engagements like for
missile seekers.
For satellite imagery, the actual impetus was the program of unmanned planetary
satellite in the 1960s that transmitted images to the ground receiving stations; the low
visual quality of the received images required the development of processing
approaches to provide the images valuable. The other impetus was the Land sat
program which began in 1972 and provided repeated worldwide exposure in digital
plan. A third impetus is the continued growth of faster and more powerful computers,
peripheral device and software that are applicable for image processing [8]. Mostly,
Infrared (IR) images are used for military applications such as automatic target
identification of military relevant land targets. A precise identification of target from a
source IR and VIS image is a challenging process due to the fact that a target can have
complicated structure and can modify shape or dimension.
The objective of this work to fuse these two types of images can join the merits of
radiation details in IR images and complete texture details in VIS images. The
precise, reliable and paired descriptors of the scene in fused images make these
methods be widely used in different applications.
Related Works
DUQ et al., [1] was proposed for fusing IR and VIS images of different resolutions
and generating high-resolution images to acquire understandable and perfect fused
images. In this method, the fusion difficulty was devised as a Total Variation (TV)
minimization problem. The data reliability term limits the pixel intensity resemblance
of the down sampled fused image with respect to the IR image and the regularization
term induces the gradient resemblance of the fused image with respect to the VIS
image.
YAN et al., [2] was proposed for IR and VIS image fusion on the basis of l 2-norm. In
this method, the fusion process was formulated as l2-norm optimization problem
where the primary term computed by the l2-norm limits the fused image to have the
 [Link] / [Link]. 6(4) (2024) Page 1117 of 12

same pixel intensities as the IR image and the secondary term measured by l2-norm
forces the fused image to have the same gradient distribution as the VIS image. Also,
two weights were introduced for optimizing the l2-norm and acquiring the fused
image effectively.
DON et al., [4] was proposed by a novel hyperspectral image fusion method. In the
method, the features of the panchromatic and hyperspectral images were
simultaneously considered. The GF was applied for generating the spatial detail image
of each hyperspectral image band efficiently.
PIA19 et al., [5] was proposed by using the deep neural network. In this technique, a
Siamese CNN was applied for estimating the activity measure and automatically
generating the weight map according to the saliency of each pixel for a pair of source
images. The source images were decomposed into low- and high-frequency sub-bands
by using the 3-level wavelet transform [5]and the fused image was acquired by
restoring the wavelet images with the scaled weight maps.
ZHA20 et al., [7] was proposed based on the GAN for fusing unmatched IR and VIS
images. In this model, the GAN two-player game was used for fusing IR and VIS
images. By using this model, the corresponding IR image from a VIS image was
generated and these two images were fused together for obtaining more information.
However, the computational complexity of this model was high and also the training
was difficult compared to other deep learning models.
SHO19 et al., [6] was proposed for fusing VIS and thermal images that focuses on
generating fused images with high VIS similarity to normal RGB images when
introducing new informative details in pedestrian regions. In this method, two types of
objective functions were applied for optimization such as a similarity metric between
the RGB input and the fused output for achieving natural image appearance and an
auxiliary pedestrian detection error for defining relevant features of the human
appearance and blending them into the output.

Limitations
From the survey a precise identification of target from a source IR and VIS image is a
challenging process due to the fact that a target can have complicated structure and
can modify shape or dimension. The major challenges found and should be
considered when developing a system for target identification are follows:
• Targets on the edge of the field of sights.
• Targets that are invisible or in shadows.
• Targets that can be heard but not observed.
• Targets under less than ideal indirect fire illumination.
• Natural and man-made obstacles
 [Link] / [Link]. 6(4) (2024) Page 1118 of 12

To overcome the limitations this research work to fuse the two types of images which
can join the merits of radiation details in IR images and complete texture details in
VIS images. The precise, reliable and paired descriptors of the scene in fused images
make these techniques be widely used in different applications mention in the
proposed method.
Proposed Methodology
The proposed work consists of three stages shown in Fig.1. First work is an enhanced
version of the image fusion technique called TSD-PS with Sturdy GF, Next, TSD-PS
with Adaptive SGF (TSD-PS-ASGF) method is proposed for sharpening the output of
SGF. Third one is TSD-Deep Neural Network with ASGF (TSD-DNN-ASGF)
method is proposed that utilizes DNN instead of PC and SML schemes to construct
the saliency weighted maps of base and detail layers. In this method, a weighted-
averaging scheme is applied to fuse the base layers whereas DNN is applied to extract
the detail layer features for preserving more information. In pre-processing step, two
scale decomposition is applied for establishing them in the fused image.
TSD-PS-SGF METHOD
The proposed method has four phases such as TSD, generation of saliency of base and
detail layers with PC and SML, respectively, computing weighting maps with SGF
and two-scale image restoration.
TSD and Saliency Map Construction
First, the source images In (i,j) are decomposed by TSD method in which a
median filter is used to approximate the base layer as:

𝑋𝑛 (𝑖, 𝑗) = 𝐼𝑛 (𝑖, 𝑗) ∗ 𝜇(𝑖, 𝑗) (1)

In Eq. (1), n represents the nth source image and μ denotes the median filter.
The range of median filter is set to 35×35 to obtain the detail layer as:

𝑌𝑛 (𝑖, 𝑗) = 𝐼𝑛 (𝑖, 𝑗) − 𝑋𝑛 (𝑖, 𝑗) (2)

[Link] / [Link]. 6(4) (2024) Page 1119 of 12

IR Image VIS Image

Preprocessing using Two Scale Decomposition

IR VIS

Base Layer Detail Layer Base Layer Detail Layer

Saliency and Weighted Map Construction using SGF Methods

* TSD-PS-SGF

* TSD-PS-ASGF

* TSD-DNN-ASGF

Final Fused Base Layer Final Fused Detail Layer

Final Fused Image

Fig .1. Proposed Fusion based Target identification Method

 [Link] / [Link]. 6(4) (2024) Page 1120 of 12

Afterwards, PC technique is adopted to create saliency maps of base layer. The PC is

computed by extracting the phase details using log-Gabor filter as:
∑𝑙 𝐸𝑙 (𝑖)
𝑃𝐶 (𝑖) = ∑ (3)
𝑙 ∑𝑛 𝐴𝑛𝑙 (𝑖)+𝜀

In Eq. (3), l is the orientation, Anl(i) denotes the sum of amplitudes, El (i)
denotes the local energy and ε refers to the small constant coefficient for preventing
the zero denominator, therefore:

𝐸𝑙 (𝑖) = √𝐹 (𝑖)2 + 𝐻 (𝑖)2 (4)

𝐹 (𝑖) = ∑𝑛 𝑠𝑛𝑙 (𝑖), 𝐻 (𝑖) = ∑𝑙 ℎ𝑛𝑙 (𝑖) (5)

𝐴𝑛𝑙 (𝑖) = ∑𝑙 √𝑠𝑛𝑙 (𝑖)2 + ℎ𝑛𝑙 (𝑖)2 (6)

In Eq. (5), snl (i) and hnl (i) are outcome of the even symmetric and odd filters,
correspondingly. A large SML is used to create the saliency maps of detail layer.
Consider the location of pixel is a, thus SML at pixel a is described like:

𝑆𝑀𝐿(𝑎) = ∑𝑘=𝜔𝑎 𝑀𝐿(𝑘 )2 (7)

In this algorithm, the pixel a is used as the centroid to construct a rectangular

window ωa. If k=(i,j), then dl is the nth detail layer and step is the distance between
pixels.
Weighted Map Construction
Initially, the weighted maps are generated by evaluating the saliency maps.
Then, the GF is applied to enhance the spatial stability of weighted maps in which the
parameters are optimized based on the IRL Swith EPCG algorithm. The weighted
maps of detail layers in VIS and IR images are PVISd and PIRd determined by
comparing the saliency maps of this layer and Similarly, the weighted maps of base
layers in VIS and IR images are also computed as:
𝑏 𝑏
𝑏 1, 𝑆𝑉𝐼𝑆 > 𝑆𝐼𝑅 𝑏 𝑏
𝑃𝑉𝐼𝑆 ={ ; 𝑃𝐼𝑅 = 1 − 𝑃𝑉𝐼𝑆 (8)
0, 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Normally, the weighted maps are not properly associated with object margins which
may generate artifacts in the fused image. Hence, SGF is carried out on each weight
map PVISd,PVISb,PIRb and PIRd with equivalent source image In(i,j). Therefore, the
resultant weighted maps of base and detail layers of both VIS and IR images are as:
𝑏 𝑏 𝑏 𝑏
𝑊𝑉𝐼𝑆 = 𝐺 (𝑃𝑉𝐼𝑆 , 𝑋𝑉𝐼𝑆 , 𝑟1 , 𝜀1 ), 𝑊𝐼𝑅 = 𝐺 (𝑃𝐼𝑅 , 𝑋𝐼𝑅 , 𝑟1 , 𝜀1 ) (9)
𝑑 𝑑 𝑑 𝑑
𝑊𝑉𝐼𝑆 = 𝐺 (𝑃𝑉𝐼𝑆 , 𝑌𝑉𝐼𝑆 , 𝑟2 , 𝜀2 ), 𝑊𝐼𝑅 = 𝐺 (𝑃𝐼𝑅 , 𝑌𝐼𝑅 , 𝑟2 , 𝜀2 ) (10)
Thus, the EPCG is used for determining the residual and the corresponding weighting
maps until the convergence is satisfied.
 [Link] / [Link]. 6(4) (2024) Page 1121 of 12

Adaptive Sturdy Guided Filtering

Given a source image𝐼𝑛 (𝑖, 𝑗), the weighted maps of base (𝑃𝑏 ) and detail layers
(𝑃𝑑 ) for IR and VIS images are applied to the ASGF. Therefore, the resultant
weighted maps are:
𝑏 𝑏 𝑏 𝑏
𝑊𝑉𝐼𝑆 = 𝐺 (𝑃𝑉𝐼𝑆 , 𝑋𝑉𝐼𝑆 , 𝑟1 , 𝜀1 ), 𝑊𝐼𝑅 = 𝐺 (𝑃𝐼𝑅 , 𝑋𝐼𝑅 , 𝑟1 , 𝜀1 ) (11)
𝑑 𝑑 𝑑 𝑑
𝑊𝑉𝐼𝑆 = 𝐺 (𝑃𝑉𝐼𝑆 , 𝑌𝑉𝐼𝑆 , 𝑟2 , 𝜀2 ), 𝑊𝐼𝑅 = 𝐺 (𝑃𝐼𝑅 , 𝑌𝐼𝑅 , 𝑟2 , 𝜀2 ) (12)

Here 𝑋𝑉𝐼𝑆 and 𝑌𝑉𝐼𝑆 are base and detail layers of VIS images, respectively. In the
same manner, 𝑋𝐼𝑅 and 𝑌𝐼𝑅 are base and detail layers of IR images, respectively.
The weight 𝑤 𝐴𝑆𝐺𝐹 (𝑎, 𝑧) is computed as a reproduction of weighted maps of
base and details layers for VIS and IR images:
𝑏 𝑏) ( 𝑑 𝑑)
𝑤 𝐴𝑆𝐺𝐹 (𝑎, 𝑧) = 𝜉𝑧′ + ((𝑃𝑉𝐼𝑆 + 𝑃𝐼𝑅 ∙ 𝑃𝑉𝐼𝑆 + 𝑃𝐼𝑅 ) (13)

𝑑 𝑑
𝑑 1, 𝑆𝑉𝐼𝑆 > 𝑆𝐼𝑅 𝑑 𝑑
Where 𝑃𝑉𝐼𝑆 ={ ; 𝑃𝐼𝑅 = 1 − 𝑃𝑉𝐼𝑆 (14)
0, 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑏 𝑏
𝑏 1, 𝑆𝑉𝐼𝑆 > 𝑆𝐼𝑅 𝑏 𝑏
𝑃𝑉𝐼𝑆 ={ ; 𝑃𝐼𝑅 = 1 − 𝑃𝑉𝐼𝑆 (15)
0, 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑏 𝑑
Here, 𝑆𝑉𝐼𝑆 and 𝑆𝑉𝐼𝑆 are saliency maps of base and detail layer for VIS images.
𝑏 𝑑
Likewise, 𝑆𝐼𝑅 and 𝑆𝐼𝑅 are the saliency maps of base and detail layer for IR
images. 𝜉𝑧′ is the offset which is defined by the naive offset selection scheme as:
𝑀𝑎𝑥(𝜔𝑖,𝑗 ) − 𝐼, 𝑖𝑓 ∆′ > 0
𝜉𝑧′ = { 𝑀𝑖𝑛(𝜔𝑖,𝑗 ) − 𝐼, 𝑖𝑓 ∆′ < 0 (16)
0, 𝑖𝑓 ∆′ = 0

Where ∆′ = 𝐼 − 𝜇𝑖,𝑗
Here, ∆′ is the intensity variance between pixel 𝑧 and the mean (𝜇𝑖,𝑗 ) of the
guided image 𝐼 in Gaussian window𝜔𝑖,𝑗 .
The proposed TSD-DNN-ASGF method is briefly explained. Assume there are
𝑁 preregistered source images(𝐼𝑛 , 𝑛 ∈ {1, … , 𝑁}). Then, TSD method is
applied to decompose the source images. For each source image𝐼𝑛 , the base
layers (𝑋𝑛 ) and the detail layers (𝑌𝑛 ) are obtained by TSD. Then, the base
layers are fused by the weighted-averaging method and the detail information
is reconstructed by the DNN. Finally, the fused image 𝐹 is reconstructed by
combining the fused base 𝑋̅ and detail layers𝑌̅.
For the detail layer𝑌1 , 𝑌2 , … , 𝑌𝑛 , a DNN i.e., VGG-19 network is applied for
extracting the deep features. After that, the weighted maps are obtained by a
 [Link] / [Link]. 6(4) (2024) Page 1122 of 12

multi-layer fusion method. At last, the fused detail layer is reconstructed by

these weighted maps and the detail layer.
Assume the detail layer 𝑌𝑛 and 𝜙𝑛𝑥,𝑚 is the feature maps of 𝑛𝑡ℎ detail layer is
extracted by the 𝑥 𝑡ℎ layer where 𝑚 (𝑚 ∈ {1,2, … , 𝑀}, 𝑀 = 64 × 2𝑖−1 ) denotes
the channel number of 𝑥 𝑡ℎ layer.
𝜙𝑛𝑥,𝑚 = 𝛷𝑥 (𝑌𝑛 ) (17)
After the deep features 𝜙𝑛𝑥,𝑚 are obtained, the activity level map 𝐶𝑛𝑥 is
computed by 𝑙1 -norm and block-based average operator. The 𝑙1 -norm of
𝜙𝑛𝑥,1:𝑀 (𝑖, 𝑗) is the activity level measure of the source detail layer. Therefore,
the primary activity level map 𝐶𝑛𝑥 is obtained as:

𝐶𝑛𝑥 (𝑖, 𝑗) = ‖𝜙𝑛𝑥,1:𝑀 (𝑖, 𝑗)‖1 (18)

Then, the block-based average operator is used for computing the final activity
level map 𝐶̂𝑛𝑥 (𝑖, 𝑗) for providing the fusion method more robust to
misregistration.
∑𝑟𝛽=−𝑟 ∑𝑟𝜃=−𝑟 𝐶𝑛𝑥 (𝑖+𝛽,𝑗+𝜃)
𝐶̂𝑛𝑥 (𝑖, 𝑗) = (2𝑟+1)2
(19)

consider 𝑟 = 1. Then, the key weighted maps 𝑊𝑛𝑥 are determined by the soft-
max operator as:
𝐶̂ 𝑥 (𝑖,𝑗)
𝑊𝑛𝑥 (𝑖, 𝑗) = ∑𝑁 𝑛 ̂𝑥
(20)
𝑛=1 𝐶𝑛 (𝑖,𝑗)

The key weighted map values 𝑊𝑛𝑥 (𝑖, 𝑗) are in the range of[0,1]. The pooling
operator in VGG-network is a type of sub-sampling method. Each time this
1
operator can resize the feature maps to times of the original size where 𝑠
𝑠
denotes the stride of the pooling operator which is equal to 2. After all the
weighted maps are obtained, these are applied to ASGF to improve the
smoothness and sharpness of the fused detail layer efficiently.
Experimental Results
The existing and proposed methods are evaluated in terms of information
theory-based metric (Mutual Information (MI)), a Visual Information Fidelity
𝐴𝐵
(VIF), an image feature-based metric(𝑄 𝐹 ), an quality of weight fusion metric
(𝑄𝑊 ), an image edge-based metric (𝑄𝐸 ) and an image structure-based
metric(𝑄 𝑆𝑆𝐼𝑀 ) to ensure the effectiveness of the proposed methods.
Table 1 Numerical Analysis of Proposed Methods for TNO Image Fusion
Datasets
 [Link] / [Link]. 6(4) (2024) Page 1123 of 12

TSD-PS- TSD-PS- TSD-DNN-

Metrics/Methods NSCT-PS-GF TSD-PS-GF
SGF ASGF ASGF

𝑴𝑰
2.14 2.43 2.51 2.62 2.83
𝑨𝑩
𝑸𝑭 0.544 0.568 0.592 0.615 0.637

𝑸𝑺𝑺𝑰𝑴 0.628 0.662 0.666 0.681 0.702

𝑽𝑰𝑭
0.4740 0.4956 0.5108 0.5397 0.5516
𝑸𝑾 0.8306 0.8351 0.8474 0.8596 0.8759
𝑸𝑬 0.1006 0.1886 0.2051 0.2284 0.2499

Image 1 Image 2 Image 3

IR Image

VIS Image

NSCT-PS-GF

TSD-PS-GF
 [Link] / [Link]. 6(4) (2024) Page 1124 of 12

TSD-PS-SGF

TSD-PS-
ASGF

TSD-DNN-
ASGF

Figure 2 Fused Image Results of Proposed Method for TNO Image Fusion Dataset

Image 1 Image 2 Image 3

IR
Image

VIS
Image

NSCT
-PS-
GF

TSD-
PS-GF
 [Link] / [Link]. 6(4) (2024) Page 1125 of 12

TSD-
PS-
SGF

TSD-
PS-
ASGF

TSD-
DNN-
ASGF
Figure 3 Fused Image Results Proposed Method for TRICLOBS Image Dataset
CONCLUSION
The enhanced image fusion technique for achieving effective military target
identification is proposed that integrates the decomposition and filtering methods to
enhance the fused image visual quality by preserving the most essential details of the
source images. This research focuses on enhancing the visual quality of the fused
image by sharpening and preserving more edge details with the minimum
computational complexity. The key objective of this enhanced fusion is to use an
adaptive robust filtering and deep learning methods in decomposing the source
images, constructing the weighted maps to fuse them and get the final fused image.
References
[1] Du, Q., Xu, H., Ma, Y., Huang, J., & Fan, F. (2018). Fusing infrared and
visible images of different resolutions via total variation model. Sensors,
18(11), 3827.
[2] Yan, H., & Li, Z. (2019). Novel model for infrared and visible image fusion
based on l_2 norm. OSA Continuum, 2(11), 3076-3090.
[3] Zhang, Y., Wei, W., & Yuan, Y. (2019). Multi-focus image fusion with
alternating guided filtering. Signal, Image and Video Processing, 13(4), 727-735.
[4] Dong, W., Xiao, S., & Qu, J. (2018). Fusion of hyperspectral and
panchromatic images with guided filter. Signal, Image and Video Processing, 12(7),
1369-1376.
[5] Piao, J., Chen, Y., & Shin, H. (2019). A New Deep Learning Based Multi-
Spectral Image Fusion Method. Entropy, 21(6), 570.
[6] Shopovska, I., Jovanov, L., & Philips, W. (2019). Deep visible and thermal
image fusion for enhanced pedestrian visibility. Sensors, 19(17), 3727.
 [Link] / [Link]. 6(4) (2024) Page 1126 of 12

[7] Zhao, Y., Fu, G., Wang, H., & Zhang, S. (2020). The Fusion of Unmatched
Infrared and Visible Images Based on Generative Adversarial Networks.
Mathematical Problems in Engineering, 2020.
[8] Feng, Y., Lu, H., Bai, J., Cao, L., & Yin, H. (2020). Fully convolutional
network-based infrared and visible image fusion. Multimedia Tools and Applications,
1-14.
[9] Ge, Y., & Jing, G. (2019). Infrared and visible image fusion using multi-
resolution convolution neural network. In Proceedings of the International
Conference on Artificial Intelligence, Information Processing and Cloud Computing
(pp. 1-5).
[10] Han, X., Lv, T., Song, X., Nie, T., Liang, H., He, B., &Kuijper, A. (2019). An
adaptive two- scale image fusion of visible and infrared images. IEEE Access, 7,
56341-56352.
[11] Jinju, J., Santhi, N., Ramar, K., & Bama, B. S. (2019). Spatial frequency
discrete wavelet transform image fusion technique for remote sensing applications.
Engineering Science and Technology, an International Journal, 22(3), 715-726.
[12] Ma, J., Ma, Y., & Li, C. (2019). Infrared and visible image fusion methods
and applications: A survey. Information Fusion, 45, 153-178.
[13] Ma, J., Liang, P., Yu, W., Chen, C., Guo, X., Wu, J., & Jiang, J. (2020).
Infrared and visible image fusion via detail preserving adversarial learning.
Information Fusion, 54, 85-98.
[14] Mishra, A., Mahapatra, S., & Banerjee, S. (2017). Modified Frei-Chen
operator-based infrared and visible sensor image fusion for real-time applications.
IEEE Sensors Journal, 17(14), 4639-4646.
[15] Piao, J., Chen, Y., & Shin, H. (2019). A New Deep Learning Based Multi-
Spectral Image Fusion Method. Entropy, 21(6), 570.
[16] Precilla, A. C., George, J., Kannan, S. R., &Prabhu, R. (2018). Modified PCA
based image fusion using feature matching. International Journal of Pure and Applied
Mathematics, 119(15), 477-483.
[17] Shao, Z., &Cai, J. (2018). Remote sensing image fusion with deep
convolutional neural network. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, 11(5), 1656-1669.
[18] Shopovska, I., Jovanov, L., & Philips, W. (2019). Deep visible and thermal
image fusion for enhanced pedestrian visibility. Sensors, 19(17), 3727.
[19] Zhang, Y., Wei, W., & Yuan, Y. (2019). Multi-focus image fusion with
alternating guided filtering. Signal, Image and Video Processing, 13(4), 727-735.
[20] Zhao, Y., Fu, G., Wang, H., & Zhang, S. (2020). The Fusion of Unmatched
Infrared and Visible Images Based on Generative Adversarial Networks.
Mathematical Problems in Engineering, 2020.