- Open Access
- Authors : Sruthi P Skaria, Shandry K K, Dr. Ojus Thomas Lee
- Paper ID : IJERTCONV9IS13009
- Volume & Issue : NCREIS – 2021 (Volume 09 – Issue 13)
- Published (First Online): 02-08-2021
- ISSN (Online) : 2278-0181
- Publisher Name : IJERT
- License: This work is licensed under a Creative Commons Attribution 4.0 International License
A Study on Hyperspectral and Multispectral Image Combination or Fusion
Sruthi P Skaria
PG Scholar, Computer Science and Engineering College of Engineering Kidangoor Kottayam, Kerala, India
Shandry K K
Assistant Professor, Computer Science and Engineering College of Engineering Kidangoor, Kottayam, Kerala, India
Dr. Ojus Thomas Lee Associate Professor, Computer Science and Engineering College of Engineering Kidangoor, Kottayam, Kerala, India
AbstractHyperspectral imaging (HSI) with otherworldly high targets occasionally encounters low-space targets that can be inferred from image sensor obstacles. Image fusion is a con- venient and practical method for processing HSI space target enhancement. It can solidify HSI and multispectral image (MSI) of higher space targets with comparable environments. In the early years, various combinations of HSI and MSI calculations were all familiar to obtain high-target HSI. In any case, you have not conducted large-scale research on the recently proposed combination of HSI and MSI. They are divided into four categories, including pan-honing or pan-sharpening, frame or matrix decomposition, tensor representation, and methods based on deep convolution neural networks.
Keywords Image fusion in Hyperspectral-Multispectral images; resolution; low-resolution; high-resolution Hyperspectral imaging
Hyperspectral imaging sensors can assemble numerous or numerous powerful gatherings in wide-range unearthly groups. Since the materials consistently have unmistakable reflectance for different frequencies, hyperspectral Image (HSI) engages precise conspicuous verification of materials owing to its high unearthly goal and wide extraordinary reach. Hyperspectral imaging has found extensive applications in distant detecting , face acknowledgment , clinical conclusion , etc Nonetheless, under limitations for imaging cameras, there is a certain tradeoff for the otherworldly goal and spatial goal. From now on, HSI with a huge number of gatherings typically has a low spatial goal to ensure a high Signal to Noise Ratio (SNR). Regardless of what is for the most part anticipated, imaging sensors can get a image with a higher spatial goal anyway with few unearthly groups, containing RGB image, panchromatic image, and multispectral image(MSI).
As shown in figure 1 the merger of HSI and MSI is an important and pragmatic way to manage the improvement of hsi space objectives the combination of hsi and msi is a pixel- level image collection the combination of HSI and MSI is a pixel- level image mixing block the image combination of high-altitude warning targets can help us improve and similar materials and have been added to different tasks including the identification of the sequence of objects inconsistent positions and changes this mix can be found in high-altitude target climate perception mineral exploration plant research and tragedy warning
Fig. 1. Figure shows a HSI-MSI fusion.
IMAGE FUSION PROCEDURE
The objective of Image Combination or Fusion is to give a consolidated image that contains an association of data from various images. Figure 2 shows the significant advancement in the picture combination measure. As a rule, the enrollment structure is viewed as an improvement issue, used to mishandle pictures and lessen costs. The picture enlistment strategy is utilized to change the impression of various images comparative
with the reference image. In this procedure, various source images are utilized for setting, where the principal image is viewed as a viewpoint image, and the primary image is changed by the reference image. In highlight extraction, the brilliant spots of the recorded image are erased to make some component maps.
By utilizing a decision operator whose primary target is to name the registered images concerning pixels or include maps, a bunch of decision guides are created. Semantic equivalence got the decision or highlight maps that probably wont give to a comparable item. It is utilized to interface these guides to a typical item to perform the fusion. This cycle is repeated for the source gotten from a comparable sort of sensors. At that point, radiometric adjustment is utilized on spatially adjusted images. Subsequently, the change of highlight maps is performed on a normal scale to get the final product in a comparative portrayal design. At long last, Image Fusion com- bines the important images into one resultant image containing an upgraded clarification of the image. The primary objective of fusion is to get a more Informative fused images.
Fig. 2. Figure shows the steps in Image Fusion Process.
TECHNIQUES IN IMAGE COMBINATION
If the strategy can be described as space and repeating re- gions. The spatial strategy monitors the potential pixel gain of the information image, where various aspects of the pixels are controlled to achieve suitable results. Use the Fourier Trans- form (FT) of the image to evaluate all associated exercises, and then check the image combination or fusion technique to obtain the resulting image. Other image combination or fusion techniques include principal component analysis (PCA), intensity hue saturation (IHS) and high-pass separation, as well as the Brovey method .
Spatial Based Methodologies
The Spatial based procedure is a straightforward image fusion strategy comprise of Max-Min, Minimum, Maximum, Simple Average and Simple Block Replace strategies .
Simple Average: It is a fusion strategy used to consoli- dated images by averaging the pixels. This method focused on all regions of the image and assuming the images are taken from a similar kind of sensor, it functions admirably . On the off chance that the images have high brightness and high contrast then it will deliver great results.
Minimum Methodologies: It chooses the most minimal power value of the pixels from images and produced a fused image . It is utilized for darker images .
MaxMin Methodologies: It chooses the averaging up- sides of the pixels littlest and biggest from the whole source images and delivered the resultant merged image.
Intensity Hue Saturation: It is a fundamental combina- tion shading procedure that changed over the RedGreenBlue images into HIS parts and subsequently force levels are parceled with panchromatic (PAN) image. Spatial contains force data and ghastly contains both shade and inundation Information of the picture. It acts in gatherings and has three multispectral bunches RedGreenBlue (RGB) of low goal. In the long run, the opposite change is performed to change over the HIS space to the principal RGB space for yielding melded images . It is an amazingly clear technique to join the images incorporates and gives a high spatial quality image. In far away distinguishing images it gives the best result and the huge drawback is that it included only three gatherings .
Principal Component Analysis: This quantifiable methodology subject to symmetric variations, which changes a great deal of impression of conceivable relating factors into a lot of head segments of clearly separated components. The standard burdens of PCA are exceptional spooky defilement and concealing contorting .
These strategies rotted the multiscale coefficients from the data images. Spatial twisting can be dealt with by the repeat or recurrence method.
Laplacian Pyramid Fusion Technique: It utilizes the interpolation sequence and Gaussian pyrami for multi- resolution examination for image fusion. Saleem et al. have detailed an improved image fusion method utilizing a dif- ferentiation pyramid change on multi-source pictures . However, it is ensured by the downside of extraction capacity which can be overwhelmed by multi-scale decay. Further, Li et al. improved the angle pyramid multi-source IF strategy which achieves a high band coefficient with the assistance of an inclination-bearing administrator .
Curvelet Transform Method: Over time, SWT has a major brand name: Relapse. For mild formulas, you can get good results. The second Curvelet is another multi-scale change; it breaks the weight of the wavelet process by leaning towards the title of the image break point.
Discrete Transform Combination or Fusion Method
: Combination or Fusion dependent on discrete movements utilizes composite pictures. In the first place, if the image is shaded, the RGB (red, green, and blue) portions of the different images are detached thusly, discrete changes are applied to the image, and afterward the normals of countless pictures are recorded and applied backward to finish the progressions to acquire a consolidated image. Contrasted and other combination or fusion techniques, (for example, Laplace pyramid methodology, Curvelet change system, and so on), DWT (Discrete Wavelet Transform) is actually prevalent. .
METHODS IN HSI-MSI COMBINATION
It can be seen from the quality of HSI-MSI fusion meth- ods that they can be divided into four categories, namely methods based on pan-honing or pan-sharpening, frame or matrix factorization, tensor representation methods, and deep convolutional neural network methods.
Pan-sharpening based HSI-MSI Combination or Fusion Approaches
The primary spatial range combination or fusion innovation expects to join low-goal MSI with high-resolution panchro- matic images(PAN) [11, 12], which is called skillet honing. Dish honing has significant and broad applications in distant detecting. There are two kinds of all encompassing crumbling strategy specialists, including CS and MRA. Composing  gives a solid audit of the all encompassing honing innovation. CS moves to the first and primary climbing illustration of low- goal MSI, which has a similar spatial size as the PAN picture, and afterward isolates the spatial information and the ghostly information of the investigated MSI into explicit sections that rely upon the change. obviously. Hence, the spatial information is mimicked by the PAN image, and the reenacted spatial information and powerful information are brought to the image region by switch moving to acquire the combined MSI. In order to reduce the distortion caused by this hybrid strategy, the histogram of the PAN image and the comparison part are coordinated before replacement. Therefore, the PAN coordination histogram has a similar mean and jitter to the modulus to be replaced. The commissioned work in this series includes intensity hue saturation (IHS), principal component analysis (PCA) and Gram-Schmidt (GS), which are in sharp
contrast to the changes in fusion technology.
Since pan-honing or sharpening is a particular instance of the HSI-MSI combination or fusion, various primers have been made to grow the skillet honing approaches for intertwining HSI and MSI. To change the container honing procedure to HSI-MSI combination or fusion, Chen et al.  first thing allotment the ridiculous gatherings of HSI a few get-togethers reliant upon the unearthly groups, and thereafter join each band of MSI with the contrasting phantom gatherings in HSI by making using of the current pan-honing or sharpening approaches. Moreover, Selva et al.  coordinate the image of high spatial goal for each phantom band of HSI through
straight backslide on high-resolution MSI, and circuit each ghastly band of the HSI with the fused high-resolution image using the pan-sharping system. Combination or fusion results watch that the fused high-resolution image can procure very much wanted combination results over a picked band in MSI for the combination. The skillet honing based HSI-MSI com- bination or fusion systems consistently have low estimation costs and can be completed rapidly. Regardless, they regularly produce remarkable fusion when the spatial resolution of HSIand MSI change sign.
MF based HSI-MSI Combination or Fusion Approaches
The Matrix Factorization based HSI-MSI fusion approaches unfold the three-dimensional fused HSI Y with the spectral mode, and acquire the network
Y(3) RSÃ—W H
The approach in this category assume that
can be decomposed as spectral basis
multiplied by coefficients
A RLÃ—W H
Both high-goal MSI and low-goal HSI can be viewed as sub- tested renditions of consolidated HSI. Indeed, the extraordinary reason D addresses high-goal HSI phantom in- formation. In light of the best strategy for showing phantom premises, MF-based strategies can generally be assigned to inadequate portrayal procedures and low-position techniques. The inadequate delivering innovation regards the otherworldly reason as an over-complete word reference and cant get shortage. They acknowledge that each extraordinary mark is an immediate combination of a couple of particles in the word reference. Word references are generally gotten from low-goal HSI through meager word reference learning computations, like K-SVD, online word reference learning, and non-negative learning. Regularly assessed by helpless coding estimations. In light of the low-position technique, the unearthly markers can be handled by low-dimensional subspace, and D is the low- position matrix. The unprecedented low D position premise generally comes from the low goal utilizing vertex segment investigation (VCA) . The combination technique dependent on network factorization intends to quantify the range premises and coefficients by taking care of more improved issues. As per the high level meaning of range premises and coeffi- cients, combination systems dependent on lattice factorization are mostly partitioned into three classifications. The primary family strategy thinks about that ghastly information and spatial information essentially depend on low-goal HSI and high-goal MSI, individually. The innovation in the following exercise accepts that low-goal HSI additionally contains spatial information, which can add to the An assessment coefficient. Generally, they initially ascertain the otherworldly reason of
the prominent HSI, and afterward compute the coefficients of the two pictures. Given the most extreme back likelihood (MAP), we characterize the computation of An as a limit comprising of a regularization term and two quadratic data change terms. The powerful reason D is frequently acquired from low goal utilizing vertex segment investigation (VCA) . The combination technique dependent on grid factorization intends to quantify the range premises and coefficients by taking care of more improved issues. As indicated by the high level meaning of range premises and coefficients, combination systems dependent on grid factorization are primarily isolated into three classifications. The principle family technique thinks about that otherworldly information and spatial information principally depend on low- goal HSI and high-goal MSI, in- dividually. The innovation in the following exercise accepts that low-goal HSI likewise contains spatial information, which can add to the An assessment coefficient. For the most part, they initially figure the unearthly reason of the prominent HSI, and afterward ascertain the coefficients of the two pictures. Because of the back greatest worth (MAP), we characterize the computation of An as a limit comprising of a regularization term and two quadratic data change terms.
Tensor Representation based HSI-MSI Combination or Fusion Approaches
The HSIs and MSIs have three estimations, and conse- quetly can be imparted by a three-dimensional tensor. Consid- ering this reality, TR has been a working subject for HSI-MSI combination or fusion. The procedures in this class rely upon different sorts of tensor Representation methodologies
Tucker decomposition  is one of the widely used TR strategies, which attenuates the high-dimensional tensor into a factor matrix and central tensor for each metric. Tucker decomposition can isolate the data of each measurement in each tensor, and then it can also establish the correlation between the data of each measurement through the central tensor. Taking advantage of citations, Tuckers decomposi- tion effectively focuses on completion, visual tracking, object recognition, and compression detection.
Tensor-train decomposition is a famous TR technique, and it characterizes another tensor rank, called tensor-train rank, which is comprised of positions of frameworks by collapsing tensor along stages of modes. Dian et al.  present a low tensor-train rank regularized HSI-MSI fusion approach. They right off the bat bunch comparable full-band patches to shape a four-dimensional tensor and afterward give the casual low tensor-train rank limitation to the four-dimensional tensor to utilize the non-local spatial-spectral similarities.
Deep CNN based HSI-MSI Combination or Fusion
Recently, due to its efficient and promising execution power, CNN has received more and more attention in many image processing applications. CNN relies on this information and can fully capture different feature images from the prepared information. Dong et al.  Suddenly proposed a deep CNN called SRCNN for super-resolution of a single image, which can achieve excellent or unparalleled presentation. The development of pan-sharpening strategy based on deep CNN
is introduced to match the noticed MSI with high- resolution panchromatic images to obtain high-resolution MSI. By changing the number of channels in the first and last convo- lutional layers, CNN-based deep pan-sharpening strategy can be easily used for HSI-MSI fusion. The CNN-based HSI-MSI deep fusion method plan is familiar with the accompanying indirect planning function g,
P = g(Q, R, ) (1)
Deep combination or fusion strategies dependent on CNN can be separated into single-branch CNN-based combination or fusion techniques and double branch CNN strategies. The branch-based CNN combination or fusion strategy initially coordinates the features of high-resolution MSI and low- resolution HSI, and afterward takes care of them to the expanded CNN to design high-resolution HSI.
As the HSI-MSI combination or fusion gets closer and closer, the ideal high-resolution HSI is dark, so it is difficult to directly evaluate the nature of the fused image. To solve this problem, two evaluation systems have been proposed. The former regards accessible HSI as an ideal fusion HSI, and simulates high-resolution MSI and low-resolution HSI by using down-sampling and down-sampling, respectively. In this sense, the HSI-MSI combination or fusion method can be applied to analog information to obtain fused HSI. Therefore, the properties of fused HSI can be assessed by relying on quality measurements of accessible reference images. The decision to reduce the resolution channel is critical to the simulation. The space downstream inspection channel must be coordinated with the PSF of the HSI image sensor. Spatial destruction techniques are usually simulated by first applying a Gaussian channel with zero mean and then leading a unified subtest. Therefore, evaluation by visual inspection is also an important advancement in identifying spatial roles and spectral distortions in the merged image.
PSNR: The peak signal to noise ratio (PSNR) is an exceptionally well-known quality measurement, for two grayscaling pictures R and S
The TR method heralds brilliant fusion results in reproduc-ing data fusion. In any case, its computational cost is still high
PSNR(R, S) = 20 log
in techniques based on contrast and pan-honing or sharpening. They also need to accurately measure the PSF and SRF of the sensor.
in which the root mean square error (RMSE) is expressed as
RMSE(R, S) = I/ R S I//P (3)
in which P represents the number of pixels. The PSNR for HSI is defined as average value of all bands. The bigger value of PSNR means better fusion result.
ERGAS: The relative dimensionless global error in synthesis (ERGAS) is represented as
Big Spatial Resolution Differences in HSI-MSI com- bination or Fusion : In the Pan-honing or sharpening, the spatial downsampling factor between PAN image and MSI factor is regularly 4. Nevertheless, the spatial downsampling factor for HSI and MSI is consistently significantly higher than
4. For example, GF-2 can secure MSI of 4m GSD, and GF-5
ERGAS(E, E) = 100/s 1/S
get MSI of 30m GSD. Right when the spatial goal of HSI and MSI have colossal differences, the substance of HSI-MSI combination or fusion is dealing with a genuinely ineffectively introduced issue, since by far most of the spatial information
in which s is spatially sub-sampling factor, and Âµ represents
the mean value of the image. The smaller ERGAS, the better the fusion results.
SAM: The Spectral Angle Mapper is an exceptionally urgent file to assess the unearthly bends, which is characterized as
is lost. Subsequently, the melded HSI may contain outrageous spatial turns. Thus, the HSI-MSI combination or fusion for gigantic spatial goal contrasts is a very troublesome issue, and more undertakings ought to be made to handle this issue. The best approach to dealing with this issue is by surveying the spatial
defilement model absolutely, which can reduce the
SAM (X, X ) = 1/M
Â§ (X j
arccos Â· XJ )
j=1 I/ Xj I/2 I/ X I/2
Zero-short Learning of HSI-MSI Fusion: Fusion tech-
nology based on deep learning usually adopts an end-to-end approach to cover planning from low-resolution HSI and high-
in which S is the number of spectral pixels, and Â· denotes inner product of two vectors. A smaller value of SAM means fewer spectral distortions.
UIQI: To beat the a few burdens of RMSE, Wang et al.  writing a list called as Universal Image Quality Index (UIQI) or Q list. The UIQI between two pictures is determined as the normal worth of all picture patches, and the UIQI for two picture fixes c and d is characterized as
resolution MSI to fusion HSI. In terms of combining the nature of HSI and the feasibility of calculation, strategies based on deep learning can achieve promising execution. However, when applied to the fusion of real information, this classifica- tion strategy has two fundamental obstacles. First, they expe- rienced the harmful effects of lack of information preparation. Gradually, HSI-MSI fusion preparation information cannot be accessed regularly. In addition, deep CNN prepared in various ways may limit the speculation ability, because the observation
Q(c, d) =
4ÂµcÂµd 2c Â· d ,
model, spectral range, and number of otherworldly groups of
Âµ2 Âµ 2 c 2 d 2
in which sigma and mu represent the variance and mean, respectively. UIQI is the average value of all spectral bands for HSI. The larger value of UIQI defines that it gives better fusion results.
OPEN PROBLEMS AND NEW GUIDELINE FOR HSI-MSI COMBINATION OR FUSION
The combination or fusion of HSI-MSI has gained huge headway in the previous decade. Regardless, there are still a few difficulties. Present recent concerns and new titles of HSI- MSI combination or fusion.
Multi-temporal Images of HSI-MSI combination or Fu- sion : The remote recognition of comparable envionments by HSI and MSI is consistent on different occasions, and different features can be realized in HSI and MSI. In the current situation, the combination or fusion of HSI-MSI is a surprising problem, because the recognition models of HSI and MSI are difficult to configure, and the image action process is also ugly. The current combination strategy rarely considers this huge and problematic issue, which requires further thinking. Although the current multi-stream image combination method is applied, the combined image has the defect of light and dark in the space where the ground object moves. One possible way to solve this problem is to find the changed location in the MSI and fade the changed area.
prepared information and test information may be unique. A feasible way to solve this problem is zero-short learning, which prepares a deep CNN from the information to be merged. In the setup system, the setup information is generated by spatially reducing the labeled HSI and MSI, and the labeled HSI is used as the deep CNN throughput. In this sense, the spatial resolution reduction technology has an impact on the preparation of the generated information, and the spatial resolution reduction system method should be mastered by paying attention to the spatial degradation between HSI and MSI.
4) Productiveness in Computation: Given the scale of development of HSI and MSI information, computing power is an important archive of HSI. In the remote detection of HSI-MSI fusion, the space size is usually very large, and the number of spectral bands usually exceeds 100. Therefore, the fusion method with low computational cost is very popular. Most matrix factorization strategies and tensor representa- tion techniques have experienced the deleterious effects of high computational complexity because they require iterative processing of complex rationalization problems. A feasible method to reduce computational overhead is low-dimensional subspace rendering, which can substantially reduce the super- natural mode by abusing the redundancy in the supernatural mode. The other method is based on deep CNN, which uses an end-to-end method to predict the combined HSI without
emphasis. In addition, they can be significantly accelerated bythe artwork processing unit.
Recently, image combination or fusion space has attracted more thinking. In this work, different image fusion strategies and their advantages and disadvantages, as well as various state-of-the-art technologies have been examined. Difficulties in various applications such as medical imaging, remote sensing, photography, and surveillance images have been dis- cussed. Finally, we discussed unique evaluation measures for image fusion strategies with or without reference. Therefore, the conclusion drawn from the overview is that each image fusion method is designed for a specific application and can be used in different combinations to obtain the best results. In the future, new image fusion technologies based on deep neural networks will be created for different fields to improve the effectiveness of the fusion method when the same processing unit is used for calculation.
Bavachan B, Krishnan DP , A survey on image fusion techniques, (2014) IJRCCT 3(3):049052
Singh N, Tanwar, P ; Image fusion using improved contourlet transform technique; Int J Recent Technol Eng (IJRTE), vol 1, no. 2 (2012)
Bai L, Xu C, Wang C; A review of fusion methods of multi-spectral image. Optik-Int J Light Electron Optics; Proceedings of the IEEE Conference 126(24)(2015) :48044807
L.-J. Deng, M. Feng, X.-C. Tai ; The fusion of panchromatic and mul- tispectral remote sensing 580 images via tensor-based sparse modeling and hyper-laplacian prior; Information Fusion 52 (2019) 7689.
J. Ma, W. Yu, C. Chen, P. Liang, X. Guo, J. Jiang, Pan-gan ; An unsupervised pan-sharpening method for remote sensing image fusion; Inf. Fus. 62 (2020) 110120.
Vivone, L. Alparone, J. Chanussot, M. Dalla Mura, A. Garzelli, G.
Licciardi, R. Restaino,L. Wald ; A critical comparison among pansharpening algorithms; IEEE Trans. Geosci. Remote 570 Sens. 53 (5) (2015) 25652586.
Z. Chen, H. Pu, B. Wang, G. Jiang ; Fusion of hyperspectral and multispectral images: A novel framework based on generalization of pan-sharpening methods; IEEE Geosci. Remote Sens. Lett. 11 (8) (2014) 14181422.
M. Selva, B. Aiazzi, F. Butera, L. Chiarantini, S. Baronti ; Hyper- sharpening: A first approach on 615 sim-ga data; IEEE J. Sel. Top.
Appl. Earth Observ. Remote Sens. 8 (6) (2015) 30083024 ,
J. M. Nascimento J. M. Dias ; Vertex component analysis: A fast
J. Bioucas-Dias, A. Plaza, G. Camps-Valls, P. Scheunders, N. Nasrabadi,
J. Chanussot ; Hyper-spectral remote sensing data analysis and future challenges; , IEEE Geosci. Remote Sens. Mag.1 (2) (2013) 636
Z. Pan, G. Healey, M. Prasad, B. Tromberg ; Face recognition in hyperspectral images ; IEEE Trans.530 Pattern Anal. Mach. Intell. 25 (12) (2003) 15521560.
H. Akbari, Y. Kosugi, K. Kojima, N. Tanaka ; Detection and analysis of the intestinal ischemia using visible and invisible hyperspectral imaging
; IEEE Trans. Biomed. Eng. 57 (8) (2010) 2011 2017.
Morris C, Rajesh RS ; Survey of spatial domain image fusion techniques
; Int JAdv Res Comput Sci Eng Inf Technol 2(3) (2014):249254.
Jasiunas MD, Kearney DA, Hopf J, Wigley GB ; Image fusion for uninhabited airborne vehicles ; (FPT). Proceedings, p 348351. (2002) IEEE.
Dong J, Dafang Z, Yaohuan H, Jinying F ; Survey of multispectral image fusion techniques in remote sensing applications; I Alcorn State University, USA (2011)
S Banu RS ; Medical image fusion by the analysis of pixel level multi- sensor using discrete wavelet Transform; In: Proceedings of the national conference on emerging trends in computing science, p 29129 (2011).
algorithm to unmix hyperspectral data IEEE Trans. Geosci. Remote Sens. 43 (4) (2005) 898910.
L. Tucker ; Some mathematical notes on three-mode factor analysis
Psychometrika 23 (8) (1996) 33363351.
R. Dian, S. Li, L. Fang ; Learning a low tensor-train rank representation for hyperspectral image super-resolution IEEE Trans. Neural Netw. Learn. Syst. 30 (9) (2019) 26722683.
Yang J, Ma Y, Yao W, Lu WT ; A spatial domain and frequency domain integrated approach to fusion multifocus images. The International archives of the photogrammetry, remote sensing and spatial Inf sciences, 37 (2008) .
Dogra A, Goyal B, Agrawal S ; From multi-scale decomposition to non- multi-scale decomposition methods: a comprehensive survey of image fusion techniques and its applications. IEEE Access 5:1604016067 (2017) .
Wu D, Yang A, Zhu L, Zhang C; Survey of multi-sensor image fusion. International conference on life system modeling and simulation. Springer, Berlin, pp 358367 (2014)
Z. Wang, A. Bovik ; A universal image quality index, IEEE Signal Process IEEE Signal Process. Lett. 9 (3) (2002)
C. Dong, C. C. Loy, K. He, X. Tang ; Learning a deep convolutional network for image superresolution in: Euro. conf. comput. vis., Springer, 2014, pp. 184199.
M. Giuseppe, C. Davide, V. Luisa, S. Giuseppe ; Pansharpening by convolutional neural networks Remote Sens. 8 (7) (2016) 594