Infrared and Visible Image Fusion Based on Implicit Neural Representations

Chang Liu; Jun Qiu; Ligen Shi; Lina Wu; Shuchen Sun

REVIEW 5 major objections 5 minor 20 references

Infrared and Visible Image Fusion Based on Implicit Neural Representations

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that infrared–visible fusion can be done by fitting a continuous coordinate-to-pixel function per image pair, without any training dataset, and that this beats nine existing methods on standard benchmarks.

desk verdict SIREN-based fusion is a sensible idea, but the SOTA claim rests on an unfair comparison and metric circularity; not ready as is. read the letter →

arxiv 2506.16773 v1 pith:LCZ6GYSQ submitted 2025-06-20 cs.CV

Shuchen Sun , Ligen Shi , Chang Liu , Lina Wu , Jun Qiu This is my paper · ORCID

classification cs.CV

keywords infraredandvisibleimagefusionimplicitneuralrepresentationSIRENcoordinate-basedMLPunsupervisedmulti-resolutionsuper-resolutionreconstruction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper recasts infrared and visible image fusion as fitting a continuous function from spatial coordinates to fused pixel values, rather than combining discrete pixel arrays or explicit features. The proposed INRFuse network—a multilayer perceptron with sine activations that is optimized on each image pair—needs no training dataset and, the authors report, outperforms nine existing fusion methods on most of six standard metrics on the TNO and RoadScene datasets. Because the learned representation is continuous, the same framework can fuse images of different resolutions and produce super-resolution reconstructions by querying denser coordinates. The central claim is that implicit neural representations are a credible, training-free alternative to conventional pixel- and feature-level fusion pipelines.

What carries the argument

The central object is the coordinate-based implicit neural representation: a multilayer perceptron with sinusoidal activation functions (SIREN) that maps normalized spatial coordinates $(x,y)$ to fused image pixel values. This continuous mapping is what decouples the fused image from a discrete pixel grid; the SIREN activation is chosen because sine functions naturally represent oscillatory signals like electromagnetic waves, and the network is regularized by a multi-term loss combining $L^2$ pixel consistency, $L^2$ gradient consistency, and an $L^1$ sparsity term on the fused gradient.

What would settle it

Fit a single INRFuse network to all 20 TNO pairs together and test it on held-out pairs; if the per-pair-optimized version no longer beats DenseFuse and U2Fusion on SCD, EN, and MS-SSIM, then the reported advantage depends on per-image optimization rather than on the implicit representation itself.

Watch

Extended reading notes

Core claim

INRFuse parameterizes the fused image as an implicit function $F_{\theta}: X \mapsto I_F(X)$ that maps normalized spatial coordinates directly to pixel values. Using a five-layer SIREN-style MLP with sine activations, it jointly optimizes pixel consistency, gradient consistency, and sparse regularization against both the infrared and visible source images, thereby preserving thermal radiation while retaining visible texture. On 20 pairs from TNO it achieves the best scores in five of six metrics (SD, VIF, SCD, EN, MS-SSIM), and on 40 pairs from RoadScene it achieves the best scores in four of six (VIF, SCD, EN, MS-SSIM), with near-best performance on the remaining metrics. The paper also demonstrates cross-resolution fusion and super-resolution reconstruction by sampling the same continuous function at varying coordinate densities.

Load-bearing premise

It is assumed that comparing per-image optimized outputs with models trained once on large datasets is a fair test, and that the hand-designed loss in Eq. (3) is a valid proxy for fusion quality.

Editorial extensions

If this is right

Fusion becomes a per-pair optimization problem: for a new infrared-visible pair, INRFuse needs only the two images and no curated training corpus.
Because the learned function is continuous, the fused image can be sampled at arbitrary resolutions, so differently sized source images can be fused directly and the output can be upsampled by querying high-density coordinates.
The same coordinate-based loss could be applied to other multimodal inputs, making INRFuse a general template for aligning and fusing signals that share spatial coordinates.
The reported metric gains (best SCD, EN, and MS-SSIM on both datasets) suggest that continuous representations preserve more source information and structure than discrete feature-combination pipelines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The comparison is asymmetric: INRFuse is optimized on each test pair, while the nine baselines are pretrained generic models; a fairer test would train a baseline on the same pairs or evaluate INRFuse with a single network fitted to all pairs and tested on unseen ones.
The per-pair optimization cost is not reported in detail; for real-time or large-batch applications the practical advantage would depend on whether this fitting cost is acceptable.
The loss in Eq. (3) could in principle be satisfied by copying the infrared image plus some visible gradient structure; an ablation removing each loss term would show whether the proposed combination is necessary for the metric gains.
The resolution-independence claim suggests a direct quantitative test: fuse images downsampled by different factors and compare the output against a full-resolution fusion using standard super-resolution metrics, not just qualitative examples.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing to know: the paper applies SIREN to infrared-visible fusion, and the evaluation that claims state-of-the-art results does not hold up. The method fits a 5-layer MLP per image pair by minimizing a loss that punishes pixel and gradient differences from both sources. That is a fine starting point, and the three-term loss is reasonable. The resolution-independent and super-resolution parts are a nice property of INR, but they are only shown qualitatively.

The real problem is the comparison. All nine baselines are trained once on large datasets and evaluated on the test pairs without adaptation. INRFuse is optimized individually on each test pair. That is not a fair comparison, and the paper offers no per-image baseline that uses the same loss. Worse, three of the six reported metrics (SCD, MS-SSIM, VIF) directly measure agreement with the source images, and EN/SD reward favorable statistics. Minimizing Eq. (3) moves the fused image toward those exact quantities. With 20 and 40 pairs and no error bars or significance tests, the reported margins are fully compatible with circularity. A simple positive control — e.g., a weighted average or a smooth interpolation optimized with the same objective — would tell the reader whether the network contributes anything beyond fitting the loss.

There are smaller issues: the paper cites neither prior INR-based fusion work nor any per-image optimization baseline, and the super-resolution demo lacks quantitative assessment. These are minor relative to the evaluation problem, but they matter for the novelty claim.

What the paper does well: it is clearly written, the method is described precisely, and the loss design is sensible. If the authors redo the evaluation with a matched per-image baseline, report error bars, and include the positive control, the core idea could be a useful addition to the fusion toolbox. As submitted, the central claim is not established.

My recommendation: engage with the paper, but only under a strict revision plan. A serious reviewer should ask for a new experimental section with per-image optimization baselines and significance testing. I would tell the editor to send it to review, with the explicit instruction that the experimental comparison needs to be rebuilt.

Referee Report

5 major / 5 minor

Summary. The paper proposes INRFuse, a method for infrared/visible image fusion that fits a SIREN MLP to each individual image pair. The network maps normalized spatial coordinates to fused pixel values and is optimized with a loss (Eq. 3) combining L2 pixel differences to both sources, L2 gradient differences to both sources, and an L1 gradient regularizer. The authors compare against nine fusion methods on the TNO (20 pairs) and RoadScene (40 pairs) datasets, reporting best scores in five of six metrics on TNO and four of six on RoadScene. They also present qualitative multi-resolution fusion and super-resolution examples. The central claim is that INRFuse outperforms existing methods while requiring no training dataset.

Significance. If the empirical claims were supported, the resolution-independent representation would be a useful property for fusing inputs from sensors with different spatial resolutions, and the per-pair optimization would be a practical alternative to large-scale training. However, the current experimental protocol does not establish the claimed superiority: the method is fitted individually to each test pair by minimizing a loss that directly rewards similarity to the source images, while all comparators are evaluated as fixed models without per-image adaptation. No error bars, significance tests, ablation studies, or non-neural per-image baselines are provided, and the multi-resolution/super-resolution section is only qualitative. The underlying idea is plausible, but the reported results are not a valid basis for the SOTA claim.

major comments (5)

[Section III-C, Tables I and II] The comparison is unfair. INRFuse is optimized per test pair against Eq. (3), whereas the nine comparator methods are generic models evaluated without any per-image adaptation. This confounds fusion quality with per-image overfitting and invalidates the reported margins as evidence of superiority. A positive control is needed: a simple per-image fusion rule (e.g., a weighted average or a hand-crafted optimization) fitted with the same loss should be compared under identical conditions.
[Section III-B/III-C and Eq. (3)] The evaluation is circular. The loss in Eq. (3) minimizes L2 pixel and gradient differences between the fused image and both source images, and the reported metrics include CC, SCD, MS-SSIM, and VIF, which measure similarity or correlation with the same source images. Optimizing the fused output against these very quantities mechanically inflates the scores, so the 'best' results in Tables I and II are partly by construction rather than evidence of a better fusion representation.
[Section III-A/III-C] No statistical support is provided for the comparisons. The tables report only mean values over 20 TNO and 40 RoadScene pairs, with no standard deviations, confidence intervals, or paired significance tests. Several reported differences (e.g., CC 0.5485 vs. 0.5495 on TNO; SD 10.072 vs. 10.162 on RoadScene) are small enough to be within run-to-run or pair-to-pair variability, so the claimed superiority is not established.
[Section III-D] The multi-resolution fusion and super-resolution claims are supported only by a qualitative figure with arbitrarily chosen scale factors (x1.2, x2.1, x3). No quantitative metrics, reference-based comparisons, or comparisons against other super-resolution or cross-resolution fusion methods are provided, so the resolution-independence advantage is not demonstrated.
[Section II-C and III-A] The method's behavior hinges on the loss weights and architecture, but no sensitivity analysis is reported. Only one value of lambda, one network width/depth, and one learning rate are used; no ablation shows how the fused image changes with these choices, making it unclear whether the reported performance is robust or tuned to the test sets.

minor comments (5)

[Section II-A] The sentence 'multi-layer perceptrons is utilized to adaptively fuse...' has a subject-verb agreement error; 'multi-layer perceptrons' should be singular or the verb should be plural.
[Section III-A] The normalization description is ambiguous: 'normalized to the range of [−1,1] with a standard deviation of 0.5' does not specify whether normalization is global per image or per batch, and a standard deviation of 0.5 is not guaranteed by that range.
[References] Reference [16] for the standard deviation metric cites a paper on fiber Bragg grating sensors, which is not an appropriate source for an image fusion evaluation metric.
[Section III-D] The resolution-independent representation is praised as avoiding 'interpolation errors', but the loss computation requires spatially aligning fused and source images, and the alignment procedure is not described; this needs clarification.
[General] The paper does not state the number of optimization iterations, convergence criterion, per-pair runtime, or whether the code will be released, which limits reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Per-image optimization against Eq. (3) inflates metrics that reward source-image similarity, so the reported SOTA is partly constructed by the evaluation protocol.

fitted input called prediction [Section II-C, Eq. (3); Section III-B; Tables I-II]
"By designing multiple loss functions, the method jointly optimizes the similarity between the fused image and the original images... L=∥I_IR(X)−I_F(X)∥2^2 + ∥I_VIS(X)−I_F(X)∥2^2 + ∥∇I_IR(X)−∇I_F(X)∥2^2 + ∥∇I_VIS(X)−∇I_F(X)∥2^2 + λ∥∇I_F(X)∥1"

The fused image I_F is produced by fitting the network separately to each test image pair, minimizing pixel and gradient L2 distances to I_IR and I_VIS. The evaluation then reports SCD, MS-SSIM, VIF, and CC, which the paper itself defines as measuring similarity/information fidelity between the fused image and those same source images. Minimizing Eq. (3) therefore directly moves the output toward high scores on the metrics used to claim superiority. The comparators are fixed generic models with no per-image adaptation, so the comparison is not matched; a non-neural per-image baseline minimizing the same objective would likely reproduce the metric pattern.

full rationale

The paper's derivation chain (INR parameterization, SIREN architecture, multi-term loss) is mathematically self-contained and contains no self-citation or imported uniqueness theorem; no step of the form 'X derives Y' is definitionally equivalent to its input. The circularity, if any, is in the empirical evaluation protocol. Eq. (3) is the objective minimized for each test pair, and the headline metrics SCD, MS-SSIM, VIF, and CC measure agreement between the fused image and the same source images used as targets in that objective; EN and SD reward favorable global statistics that the averaging/smoothing effect of the loss tends to produce. Consequently, the reported gains over nine generic methods are constructionally favored by the per-image fitting protocol, which the abstract summarizes as 'without the necessity for a training dataset.' A matched per-image baseline with the same loss is absent, and no significance tests or error bars accompany the small margins in Tables I and II. This is a partial, evaluation-level circularity: the central empirical claim reduces in part to fitting the evaluation-related objective, though the method itself is not a derivation that assumes its own conclusion. Score 6 reflects that the reported superiority is not independently established, while stopping short of claiming the mathematical derivation is circular.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method's central result depends on a few hand-selected hyperparameters (lambda, network size, learning rate) and on the assumption that the hand-crafted loss adequately represents fusion quality. No new conceptual entities are introduced.

free parameters (4)

lambda = 1.0
Sparsity regularization weight; chosen by hand without sensitivity analysis.
network width = 256
Number of neurons per layer; chosen without ablation.
network depth = 5
Number of layers; chosen without ablation.
learning rate = 0.001
Adam learning rate; no schedule mentioned.

assumptions (3)

domain assumption SIREN can represent high-frequency image signals
Relies on Sitzmann et al.'s demonstration that sinusoidal activations with proper initialization can fit complex signals; not re-derived here.
domain assumption The loss in Eq. (3) is a valid objective for fusion quality
Assumes that minimizing pixel and gradient difference to both source images yields a perceptually good fused image; no user study or perceptual validation.
domain assumption The network can encode both modalities in a single scalar output
For RGB visible and grayscale infrared, the paper does not specify how channels are handled; assumes a single intensity channel suffices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Infrared and Visible Image Fusion Based on Implicit Neural Representations." pith.science (2026). https://pith.science/paper/LCZ6GYSQ

@misc{pith2026250616773,
  author       = {Pith},
  title        = {Pith review of: Infrared and Visible Image Fusion Based on Implicit Neural Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCZ6GYSQ}},
  note         = {Machine review of arXiv:2506.16773}
}

read the original abstract

Infrared and visible light image fusion aims to combine the strengths of both modalities to generate images that are rich in information and fulfill visual or computational requirements. This paper proposes an image fusion method based on Implicit Neural Representations (INR), referred to as INRFuse. This method parameterizes a continuous function through a neural network to implicitly represent the multimodal information of the image, breaking through the traditional reliance on discrete pixels or explicit features. The normalized spatial coordinates of the infrared and visible light images serve as inputs, and multi-layer perceptrons is utilized to adaptively fuse the features of both modalities, resulting in the output of the fused image. By designing multiple loss functions, the method jointly optimizes the similarity between the fused image and the original images, effectively preserving the thermal radiation information of the infrared image while maintaining the texture details of the visible light image. Furthermore, the resolution-independent characteristic of INR allows for the direct fusion of images with varying resolutions and achieves super-resolution reconstruction through high-density coordinate queries. Experimental results indicate that INRFuse outperforms existing methods in both subjective visual quality and objective evaluation metrics, producing fused images with clear structures, natural details, and rich information without the necessity for a training dataset.

Figures

Figures reproduced from arXiv: 2506.16773 by the authors.

**Figure 2.** (a) (b) represent the fusion results of different methods on [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Fusion of Images with Varying Resolutions and Super-Resolution Reconstruction of Fused Images. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

[1]

The laplacian pyramid as a compact image code

Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. InReadings in Computer Vision, pages 671–679. Elsevier, 1987

work page 1987
[2]

Infrared and visible image fusion with convolutional neural networks.International Journal of Wavelets, Mul- tiresolution and Information Processing, 16(03):1850018, 2018

Yu Liu, Xun Chen, Juan Cheng, Hu Peng, and Zengfu Wang. Infrared and visible image fusion with convolutional neural networks.International Journal of Wavelets, Mul- tiresolution and Information Processing, 16(03):1850018, 2018

work page 2018
[3]

Learning continuous image representation with local implicit image function

Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628–8638, 2021

2021
[4]

Implicit neural representations with periodic activation functions.Advances in Neural Information Processing Systems, 33:7462–7473, 2020

Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions.Advances in Neural Information Processing Systems, 33:7462–7473, 2020

work page 2020
[5]

Tno image fusion dataset.Figshare

Alexander Toet et al. Tno image fusion dataset.Figshare. data, 4, 2014

work page 2014
[6]

U2fusion: A unified unsupervised image fusion net- work.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502–518, 2020

Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion net- work.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502–518, 2020

work page 2020
[7]

Remote sensing image fusion using the curvelet transform.Information Fusion, 8(2):143–156, 2007

Filippo Nencini, Andrea Garzelli, Stefano Baronti, and Luciano Alparone. Remote sensing image fusion using the curvelet transform.Information Fusion, 8(2):143–156, 2007

work page 2007
[8]

Mdlatlrr: A novel decomposition method for infrared and visible image fusion.IEEE Transactions on Image Processing, 29:4733– 4746, 2020

Hui Li, Xiao-Jun Wu, and Josef Kittler. Mdlatlrr: A novel decomposition method for infrared and visible image fusion.IEEE Transactions on Image Processing, 29:4733– 4746, 2020

work page 2020

Show all 20 references

[9]

Densefuse: A fusion approach to infrared and visible images.IEEE Transactions on Image Processing, 28(5):2614–2623, 2018

Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images.IEEE Transactions on Image Processing, 28(5):2614–2623, 2018

2018
[10]

Fusiongan: A generative adversarial network for infrared and visible image fusion.Information Fusion, 48:11–26, 2019

Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusiongan: A generative adversarial network for infrared and visible image fusion.Information Fusion, 48:11–26, 2019

2019
[11]

Unsupervised misaligned infrared and visible image fusion via cross-modality im- age generation and registration

D Wang, J Liu, X Fan, and R Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality im- age generation and registration. arxiv 2022.arXiv preprint arXiv:2205.11876, 2022

2022 arXiv
[12]

Datfuse: Infrared and visible image fusion via dual attention transformer.IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159–3172, 2023

Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer.IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159–3172, 2023

2023
[13]

Ppt fusion: Pyramid patch transformerfor a case study in image fusion.arXiv preprint arXiv:2107.13967, 2021

Yu Fu, TianYang Xu, XiaoJun Wu, and Josef Kittler. Ppt fusion: Pyramid patch transformerfor a case study in image fusion.arXiv preprint arXiv:2107.13967, 2021

2021 arXiv
[14]

Ydtr: Infrared and visible image fusion via y-shape dynamic transformer.IEEE Transactions on Multimedia, 25:5413–5428, 2023

Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer.IEEE Transactions on Multimedia, 25:5413–5428, 2023

2023
[15]

Assessment of image fusion procedures using entropy, image quality, and multispectral classification

J Wesley Roberts, Jan A Van Aardt, and Fethi Babikker Ahmed. Assessment of image fusion procedures using entropy, image quality, and multispectral classification. Journal of Applied Remote Sensing, 2(1):023522, 2008

2008
[16]

In-fibre bragg grating sensors.Measurement Science and Technology, 8(4):355, 1997

Yun-Jiang Rao. In-fibre bragg grating sensors.Measurement Science and Technology, 8(4):355, 1997

1997
[17]

Infrared and visible im- age fusion methods and applications: A survey.Information Fusion, 45:153–178, 2019

Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible im- age fusion methods and applications: A survey.Information Fusion, 45:153–178, 2019

2019
[18]

A new image quality metric for image fusion: The sum of the correlations of differences.Aeu-international Journal of Electronics and Communications, 69(12):1890–1896, 2015

V Aslantas and Emre Bendes. A new image quality metric for image fusion: The sum of the correlations of differences.Aeu-international Journal of Electronics and Communications, 69(12):1890–1896, 2015

2015
[19]

Perceptual quality assessment for multi-exposure image fusion.IEEE Trans- actions on Image Processing, 24(11):3345–3356, 2015

Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion.IEEE Trans- actions on Image Processing, 24(11):3345–3356, 2015

2015
[20]

A new image fusion performance metric based on visual information fidelity.Information Fusion, 14(2):127–135, 2013

Yu Han, Yunze Cai, Yin Cao, and Xiaoming Xu. A new image fusion performance metric based on visual information fidelity.Information Fusion, 14(2):127–135, 2013

2013

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

The laplacian pyramid as a compact image code

Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. InReadings in Computer Vision, pages 671–679. Elsevier, 1987

work page 1987

[2] [2]

Infrared and visible image fusion with convolutional neural networks.International Journal of Wavelets, Mul- tiresolution and Information Processing, 16(03):1850018, 2018

Yu Liu, Xun Chen, Juan Cheng, Hu Peng, and Zengfu Wang. Infrared and visible image fusion with convolutional neural networks.International Journal of Wavelets, Mul- tiresolution and Information Processing, 16(03):1850018, 2018

work page 2018

[3] [3]

Learning continuous image representation with local implicit image function

Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628–8638, 2021

2021

[4] [4]

Implicit neural representations with periodic activation functions.Advances in Neural Information Processing Systems, 33:7462–7473, 2020

Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions.Advances in Neural Information Processing Systems, 33:7462–7473, 2020

work page 2020

[5] [5]

Tno image fusion dataset.Figshare

Alexander Toet et al. Tno image fusion dataset.Figshare. data, 4, 2014

work page 2014

[6] [6]

U2fusion: A unified unsupervised image fusion net- work.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502–518, 2020

Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion net- work.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502–518, 2020

work page 2020

[7] [7]

Remote sensing image fusion using the curvelet transform.Information Fusion, 8(2):143–156, 2007

Filippo Nencini, Andrea Garzelli, Stefano Baronti, and Luciano Alparone. Remote sensing image fusion using the curvelet transform.Information Fusion, 8(2):143–156, 2007

work page 2007

[8] [8]

Mdlatlrr: A novel decomposition method for infrared and visible image fusion.IEEE Transactions on Image Processing, 29:4733– 4746, 2020

Hui Li, Xiao-Jun Wu, and Josef Kittler. Mdlatlrr: A novel decomposition method for infrared and visible image fusion.IEEE Transactions on Image Processing, 29:4733– 4746, 2020

work page 2020