Pith. sign in

REVIEW 4 major objections 5 minor 71 references

FSATFusion: Frequency-Spatial Attention Transformer for Infrared and Visible Image Fusion

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read FSATFusion claims that combining frequency-spatial attention with an improved Transformer yields the best infrared-visible fusion quality, efficiency, and generalization.

desk verdict Solid but incremental IVIF paper: the architecture is a competent assembly of known attention components, and the 'superior' claim needs statistical support. read the letter →

arxiv 2506.10366 v1 pith:6TGG2SSE submitted 2025-06-12 cs.CV

Tianpei Zhang , Jufeng Zhao , Yiming Zhu , Guangmang Cui , Yuhan Lyu This is my paper
classification cs.CV
keywords infraredandvisibleimagefusionTransformerfrequency-spatialattentionmechanismContextBroadcastobjectdetectiongeneralization
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

The paper tries to establish that a Transformer-based fusion network, FSATFusion, can combine infrared and visible images better and faster than eleven existing methods. It argues that convolution-based and plain-Transformer fusion lose information because they ignore global context and frequency-domain cues. FSATFusion adds a frequency-spatial attention module and an improved Transformer to keep both global structure and salient thermal details. If true, the method gives sharper fused images, stronger generalization to RGB and near-infrared tasks, and better downstream object detection than current baselines.

What carries the argument

The load-bearing object is the FSAT module, a stack of an Improved Transformer Module (ITM) and a Frequency-Spatial Attention Mechanism (FSAM). ITM inserts Context Broadcast after the MLP layer to encourage the dense attention that plain Vision Transformers need for efficient learning; FSAM partitions channels into groups, assigns each group a different 2D-DCT frequency component to build a frequency attention map, then applies max and standard-deviation pooling with a 7x7 convolution to build a spatial attention map. Together these two stages are what carry the argument: they aim to preserve global frequency information and localize salient targets without losing either.

What would settle it

Re-run FSATFusion and the eleven baselines several times on TNO and MSRS with different random seeds, compute confidence intervals for MI, NCIE, Qabf, QP, Qy, VIF, and YOLOv5 mAP, and check whether the intervals overlap; the tied mAP@0.5:0.95 with DATFuse already shows the kind of gap that could vanish under repeated runs.

Watch

Extended reading notes

Core claim

The central claim is that FSATFusion outperforms eleven state-of-the-art infrared-visible fusion methods in quality, efficiency, and generalization. The network processes each image pair through a Frequency-Spatial Attention Transformer made of an Improved Transformer Module with Context Broadcast for long-range dependencies, followed by a Frequency-Spatial Attention Mechanism that splits feature channels and applies 2D-DCT frequency components together with max and standard-deviation spatial pooling to emphasize salient regions. On TNO and MSRS the paper reports first-place mean scores on five of six metrics, the fastest average runtime across four datasets, and improved YOLOv5 detection on labeled MSRS pairs.

Load-bearing premise

The claim of superiority assumes the reported average metric differences reflect real improvement rather than run-to-run noise, since the paper gives one mean per method without error bars or significance tests.

Editorial extensions

If this is right

  • Fused images from TNO and MSRS should show sharper edges, natural brightness, and more prominent infrared targets than all eleven compared methods.
  • The same trained model should fuse infrared-RGB pairs on RoadScene and near-infrared-RGB pairs on RGB-NIR without modification, retaining detail and color fidelity.
  • YOLOv5 should detect more small or motion-blurred pedestrians in MSRS fused images than with any baseline fusion method.
  • The network should run faster on average than all eleven compared methods across the four test datasets.

Reading between the lines

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

  • A consequence the paper leaves implicit is that its 2D-DCT frequency attention block could be lifted into other fusion or restoration networks, since the block is architecture-agnostic.
  • The paper's own future-work suggestion of multi-exposure and multi-focus fusion is a plausible test bed, because those tasks also combine global intensity structure with local detail.
  • The single-run averages would be more convincing with confidence intervals; without them, small gaps such as the tied detection mAP with DATFuse may not indicate a real advantage.
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.

Referee Report

4 major / 5 minor

Summary. The paper presents FSATFusion, an unsupervised end-to-end network for infrared and visible image fusion. The network uses a shallow CNN encoder followed by N frequency-spatial attention Transformer (FSAT) modules, each containing an Improved Transformer Module (ITM) with a Context Broadcast layer and a Frequency-Spatial Attention Module (FSAM) that applies 2D-DCT-based channel attention and max/std spatial attention. Training uses a weighted combination of pixel, texture, and SSIM losses. The model is trained on LLVIP and evaluated against eleven existing methods on TNO, MSRS, RoadScene, and RGB-NIR datasets with six quantitative metrics, an ablation study, a runtime comparison, and a YOLOv5 object detection experiment. The authors claim superior fusion quality, generalization, and efficiency.

Significance. The architecture is a reasonable and lightweight combination of Transformer self-attention, Context Broadcast, and frequency-domain channel attention, and the authors provide several strengths: the code is publicly released, the evaluation spans four datasets and eleven comparators, and the runtime analysis demonstrates a genuinely fast inference time. However, the paper's main claim of superiority is not currently supported by the quantitative evidence. Every comparison is a single mean without variance or significance testing, several decisive differences are extremely small, and the paper's own tables show that the method is not best on every metric or dataset. These issues are fixable but require either additional statistical evidence or suitably qualified claims.

major comments (4)
  1. [Sec. IV-C; Figs. 5, 6, 11, 12; Tabs. II, IV] The central claim that FSATFusion is 'superior' to the eleven compared methods is not statistically established: all quantitative results are reported as single means with no error bars, confidence intervals, or significance tests, and some of the decisive gaps are very small (e.g., in Tab. IV the mAP@[0.5,0.95] values for FSATFusion and DATFuse are both 0.658, and the mAP@0.65 values are 0.844 vs. 0.843). Given test sets of only 41, 81, 75, and 88 images, these differences could well be within run-to-run or sampling noise, so the statement in Sec. IV-C that the method 'consistently ranks first' is stronger than the evidence justifies.
  2. [Sec. IV-D; Tab. II] The model hyperparameters N (number of FSAT modules) and the loss weights (alpha, beta, gamma) appear to be selected using the same benchmarks on which the final comparisons are reported: the ablation study compares N=1,2,3,4 in Tab. II and chooses N=2, and the loss weights are set to (1,10,100) without a described validation split. If the ablation metrics were computed on the MSRS test set, the reported results are selection-optimistic and the advantage over the ablated baselines may be partly due to fitting the test set rather than to the design itself.
  3. [Abstract; Sec. IV-C; Sec. IV-E] The unqualified claim of 'superior fusion quality and efficiency' in the abstract is contradicted by the paper's own results: on MSRS, Q_P is slightly below SwinFusion (Sec. IV-C); on RoadScene, Q_abf is not the best (Sec. IV-E.1); and on RGB-NIR, both Q_abf and VIF are not the best (Sec. IV-E.2). The claims should be restricted to the metrics and datasets where the differences are actually favorable and statistically reliable.
  4. [Sec. III-B] The architecture description does not specify the patch size M used in the PatchEmbedding operation of ITM, nor the assignment of 2D-DCT frequency indices (a_n, b_n) to the channel groups in Eq. (13); these are free parameters that determine the actual model. The provided code may resolve the ambiguity, but the paper itself is not self-contained for reproducing the exact architecture.
minor comments (5)
  1. [Sec. III-C] Equation (17) contains a typographical error in the summation bound: 'WXw−1' should be 'W' with the sum over w=1 to W.
  2. [Abstract] The phrase 'discriminate features' should read 'discriminative features'.
  3. [Sec. IV-C] In the final paragraph of Sec. IV-C, 'FSATFsuion' is a typo for 'FSATFusion'.
  4. [Sec. III-B] The text 'Attention metrics are calculated' should read 'Attention matrices are calculated'.
  5. [References] References [2] and [46] are the same arXiv preprint and should be merged into a single entry.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed superiority rests on external benchmarks and external building blocks, with no load-bearing reduction to the paper's own inputs.

full rationale

FSATFusion's derivation chain is self-contained. The proposed modules are compositions of externally published components: the ITM is structured after Swin Transformer [52] with Context Broadcast [26], and the FSAM uses 2D-DCT frequency attention in the style of FcaNet [54] plus CBAM-style spatial attention [53]. None of these is an author-self citation, and no uniqueness theorem is imported from the authors' prior work. The quantitative claims are evaluated on fixed public datasets (TNO, MSRS, RoadScene, RGB-NIR) using six metrics (MI, NCIE, Qabf, QP, Qy, VIF) that are not the training losses. Although the loss includes an SSIM term against the pixelwise maximum of the two sources (Eqs. 19-20), the reported Qy metric is a per-source structural similarity measure and is not the same functional, so the Qy outcome is not forced by construction. The self-citations [12,13] appear only inside a broad related-work citation list and in a non-specific reference to 'previous studies'; no central argument depends on them. Hyperparameter and module-count choices are ablation-based model selection, not a fitted parameter relabeled as a prediction. Thus no circular step is present, and the correct finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard empirical machine-learning assumptions: training on LLVIP transfers to other datasets, the chosen metrics capture fusion quality, and the FSAM frequency-decomposition scheme preserves relevant information. No new physical entities are introduced. The free parameters (loss weights, N, M, DCT frequency indices) are hand-chosen or selected via ablation on the evaluation benchmarks.

free parameters (5)
  • Loss weights alpha, beta, gamma = 1, 10, 100
    Set manually in Sec. III-C/IV-A; no sensitivity analysis, only whole-term ablation, and chosen on the same benchmark sets used for final evaluation.
  • Number of FSAT modules N = 2
    Selected via ablation (N=1,2,3,4) on TNO/MSRS in Sec. IV-D, then used for all reported results; this is model selection on test data.
  • Patch size M in ITM = not reported
    ITM partitions feature maps into H/M x W/M patches (Sec. III-B) but M is never specified, making the architecture underspecified.
  • DCT frequency indices (a_n, b_n) per channel group = not reported
    FSAM assigns different 2D-DCT frequencies to N channel groups (Eq. 13) but never lists the chosen indices; the method's behavior depends on this undocumented choice.
  • Channel dimension C = 16
    Set to 16 in Sec. III-A to enable detailed feature representation; a hand-chosen design parameter.
assumptions (3)
  • domain assumption Training on LLVIP generalizes to TNO, MSRS, RoadScene, and RGB-NIR
    The model is trained only on LLVIP and tested on other datasets (Sec. IV-A); no domain-shift or validation analysis is provided.
  • domain assumption The six evaluation metrics (MI, NCIE, Qabf, QP, Qy, VIF) reflect fusion quality
    Superiority is judged entirely by these metrics (Sec. IV-B); no human evaluation or robust task-level validation beyond one small detection test.
  • ad hoc to paper Sampling one 2D-DCT coefficient per channel group preserves frequency information
    FSAM's core claim (Sec. III-B, Eqs. 9-14) is that avoiding global average pooling prevents information loss, but only N of H*W frequency components are used; this is an unsubstantiated modeling assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FSATFusion: Frequency-Spatial Attention Transformer for Infrared and Visible Image Fusion." pith.science (2026). https://pith.science/paper/6TGG2SSE

@misc{pith2026250610366,
  author       = {Pith},
  title        = {Pith review of: FSATFusion: Frequency-Spatial Attention Transformer for Infrared and Visible Image Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TGG2SSE}},
  note         = {Machine review of arXiv:2506.10366}
}
read the original abstract

The infrared and visible images fusion (IVIF) is receiving increasing attention from both the research community and industry due to its excellent results in downstream applications. Existing deep learning approaches often utilize convolutional neural networks to extract image features. However, the inherently capacity of convolution operations to capture global context can lead to information loss, thereby restricting fusion performance. To address this limitation, we propose an end-to-end fusion network named the Frequency-Spatial Attention Transformer Fusion Network (FSATFusion). The FSATFusion contains a frequency-spatial attention Transformer (FSAT) module designed to effectively capture discriminate features from source images. This FSAT module includes a frequency-spatial attention mechanism (FSAM) capable of extracting significant features from feature maps. Additionally, we propose an improved Transformer module (ITM) to enhance the ability to extract global context information of vanilla Transformer. We conducted both qualitative and quantitative comparative experiments, demonstrating the superior fusion quality and efficiency of FSATFusion compared to other state-of-the-art methods. Furthermore, our network was tested on two additional tasks without any modifications, to verify the excellent generalization capability of FSATFusion. Finally, the object detection experiment demonstrated the superiority of FSATFusion in downstream visual tasks. Our code is available at https://github.com/Lmmh058/FSATFusion.

Figures

Figures reproduced from arXiv: 2506.10366 by the authors.

Figure 1
Figure 1. Comparison of fusion results for a pair of images from MSRS dataset. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A schematic overview of the proposed FSATFusion network is presented, highlighting the central components of the FSAT module. The architecture [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Qualitative comparison of FSATFusion and eleven comparison [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: Quantitative comparison of FSATFusion and eleven comparison methods on TNO dataset. Our method is represented by the red line. Mean values of [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Quantitative comparison of FSATFusion and eleven comparison methods on MSRS dataset. Our method is represented by the red line. Mean values [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Structure of network ablation experiment. We conducted structural analysis of FSAT using three structures: (a), (b), and (c). Meanwhile, we analyzed [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of FSATFusion ablation experiment. For clearer comparison, we highlighted two important parts in each figure, marked with [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison of FSATFusion and eleven comparison [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 11
Figure 11. Figure 11: Quantitative comparison of FSATFusion and eleven comparison methods on RoadScene dataset. Our method is represented by the red line. Mean [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Quantitative comparison of FSATFusion and eleven comparison methods on RGB-NIR dataset. Our method is represented by the red line. Mean [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Detection performance of FSATFusion and eleven comparison [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 53 canonical work pages

  1. [1]

    Datransnet: dynamic attention transformer network for infrared small target detection,

    C. Hu, Y . Huang, K. Li, L. Zhang, C. Long, Y . Zhu, T. Pu, and Z. Peng, “Datransnet: dynamic attention transformer network for infrared small target detection,”IEEE Geoscience and Remote Sensing Letters, 2025

    work page 2025

  2. [3]

    Object fusion tracking based on visible and infrared images: A comprehensive review,

    X. Zhang, P. Ye, H. Leung, K. Gong, and G. Xiao, “Object fusion tracking based on visible and infrared images: A comprehensive review,” Information Fusion, vol. 63, pp. 166–187, 2020

    work page 2020

  3. [4]

    A review of object detection based on deep learning,

    Y . Xiao, Z. Tian, J. Yu, Y . Zhang, S. Liu, S. Du, and X. Lan, “A review of object detection based on deep learning,”Multimedia Tools and Applications, vol. 79, pp. 23 729–23 791, 2020

    work page 2020

  4. [5]

    Fusion method for infrared and visible images by using non-negative sparse representation,

    J. Wang, J. Peng, X. Feng, G. He, and J. Fan, “Fusion method for infrared and visible images by using non-negative sparse representation,” Infrared Physics & Technology, vol. 67, pp. 477–489, 2014

    work page 2014

  5. [6]

    Infrared and visible image fusion using multi-scale nsct and rolling-guidance filter,

    A. Selvaraj and P. Ganesan, “Infrared and visible image fusion using multi-scale nsct and rolling-guidance filter,”IET Image Processing, vol. 14, no. 16, pp. 4210–4219, 2020

    work page 2020

  6. [7]

    Densefuse: A fusion approach to infrared and visible images,

    H. Li and X.-J. Wu, “Densefuse: A fusion approach to infrared and visible images,”IEEE Transactions on Image Processing, vol. 28, no. 5, pp. 2614–2623, 2018

    2018

  7. [8]

    Stfnet: Self-supervised transformer for infrared and visible image fusion,

    Q. Liu, J. Pi, P. Gao, and D. Yuan, “Stfnet: Self-supervised transformer for infrared and visible image fusion,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 8, no. 2, pp. 1513–1526, 2024

    work page 2024

  8. [9]

    Uud-fusion: An unsupervised universal image fusion approach via generative diffusion model,

    X. Wang, L. Fang, J. Zhao, Z. Pan, H. Li, and Y . Li, “Uud-fusion: An unsupervised universal image fusion approach via generative diffusion model,”Computer Vision and Image Understanding, vol. 249, p. 104218, 2024

    work page 2024

Show all 71 references
  1. [10]

    Hbanet: A hybrid boundary- aware attention network for infrared and visible image fusion,

    X. Luo, J. Zhang, L. Wang, and D. Niu, “Hbanet: A hybrid boundary- aware attention network for infrared and visible image fusion,”Computer Vision and Image Understanding, vol. 249, p. 104161, 2024

    2024

  2. [11]

    Fusiondiff: A unified im- age fusion network based on diffusion probabilistic models,

    Z. Huang, S. Yang, J. Wu, L. Zhu, and J. Liu, “Fusiondiff: A unified im- age fusion network based on diffusion probabilistic models,”Computer Vision and Image Understanding, vol. 244, p. 104011, 2024

    2024

  3. [12]

    Exploring state space model in wavelet domain: An infrared and visible image fusion network via wavelet transform and state space model,

    T. Zhang, Y . Zhu, J. Zhao, G. Cui, and Y . Zheng, “Exploring state space model in wavelet domain: An infrared and visible image fusion network via wavelet transform and state space model,”arXiv preprint arXiv:2503.18378, 2025

    2025 arXiv

  4. [13]

    Daaf: Degradation-aware adaptive fusion framework for robust infrared and visible images fusion,

    T. Zhang, J. Zhao, Y . Zhu, G. Cui, Y . Jing, and Y . Lyu, “Daaf: Degradation-aware adaptive fusion framework for robust infrared and visible images fusion,”arXiv preprint arXiv:2504.10871, 2025

    2025

  5. [14]

    Sgfusion: A saliency guided deep- learning framework for pixel-level image fusion,

    J. Liu, R. Dian, S. Li, and H. Liu, “Sgfusion: A saliency guided deep- learning framework for pixel-level image fusion,”Information Fusion, vol. 91, pp. 205–214, 2023

    2023

  6. [15]

    U2fusion: A unified unsupervised image fusion network,

    H. Xu, J. Ma, J. Jiang, X. Guo, and H. Ling, “U2fusion: A unified unsupervised image fusion network,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 1, pp. 502–518, 2020

    2020

  7. [16]

    Rfn-nest: An end-to-end residual fusion network for infrared and visible images,

    H. Li, X.-J. Wu, and J. Kittler, “Rfn-nest: An end-to-end residual fusion network for infrared and visible images,”Information Fusion, vol. 73, pp. 72–86, 2021

    2021

  8. [17]

    Cufd: An encoder– decoder network for visible and infrared image fusion based on common and unique feature decomposition,

    H. Xu, M. Gong, X. Tian, J. Huang, and J. Ma, “Cufd: An encoder– decoder network for visible and infrared image fusion based on common and unique feature decomposition,”Computer Vision and Image Under- standing, vol. 218, p. 103407, 2022

    2022

  9. [18]

    Fusiongan: A generative adversarial network for infrared and visible image fusion,

    J. Ma, W. Yu, P. Liang, C. Li, and J. Jiang, “Fusiongan: A generative adversarial network for infrared and visible image fusion,”Information fusion, vol. 48, pp. 11–26, 2019

    2019

  10. [19]

    Apnet: Adversarial learning assistance and perceived importance fusion network for all-day rgb- t salient object detection,

    W. Zhou, Y . Zhu, J. Lei, J. Wan, and L. Yu, “Apnet: Adversarial learning assistance and perceived importance fusion network for all-day rgb- t salient object detection,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 6, no. 4, pp. 957–968, 2022

    2022

  11. [20]

    Lgabl: Uhd multi-exposure image fusion via local and global aware bilateral learning,

    D. Wang, Z. Zheng, W. Ding, and X. Jia, “Lgabl: Uhd multi-exposure image fusion via local and global aware bilateral learning,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 8, no. 2, pp. 1362–1375, 2024

    2024

  12. [21]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  13. [22]

    Ganmcc: A generative adversarial network with multiclassification constraints for infrared and visible image fusion,

    J. Ma, H. Zhang, Z. Shao, P. Liang, and H. Xu, “Ganmcc: A generative adversarial network with multiclassification constraints for infrared and visible image fusion,”IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–14, 2020

    2020

  14. [23]

    Itfuse: An interactive transformer for infrared and visible image fusion,

    W. Tang, F. He, and Y . Liu, “Itfuse: An interactive transformer for infrared and visible image fusion,”Pattern Recognition, vol. 156, p. 110822, 2024

    2024

  15. [24]

    Ydtr: Infrared and visible image fusion via y-shape dynamic transformer,

    ——, “Ydtr: Infrared and visible image fusion via y-shape dynamic transformer,”IEEE Transactions on Multimedia, vol. 25, pp. 5413–5428, 2022

    2022

  16. [25]

    Datfuse: Infrared and visible image fusion via dual attention transformer,

    W. Tang, F. He, Y . Liu, Y . Duan, and T. Si, “Datfuse: Infrared and visible image fusion via dual attention transformer,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 7, pp. 3159– 3172, 2023

    2023

  17. [26]

    Scratching visual transformer’s back with uniform attention,

    N. Hyeon-Woo, K. Yu-Ji, B. Heo, D. Han, S. J. Oh, and T.-H. Oh, “Scratching visual transformer’s back with uniform attention,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 5807–5818

    2023

  18. [27]

    The laplacian pyramid as a compact image code,

    P. J. Burt and E. H. Adelson, “The laplacian pyramid as a compact image code,” inReadings in computer vision. Elsevier, 1987, pp. 671–679

    1987

  19. [28]

    Infrared and visible images fusion method based on discrete wavelet transform,

    L. Zhan, Y . Zhuang, and L. Huang, “Infrared and visible images fusion method based on discrete wavelet transform,”Journal of Computers, vol. 28, no. 2, pp. 57–71, 2017

    2017

  20. [29]

    Infrared and visible image fusion scheme based on nsct and low-level visual features,

    H. Li, H. Qiu, Z. Yu, and Y . Zhang, “Infrared and visible image fusion scheme based on nsct and low-level visual features,”Infrared Physics & Technology, vol. 76, pp. 174–184, 2016

    2016

  21. [30]

    The infrared and visible image fusion algorithm based on target separation and sparse representation,

    X. Lu, B. Zhang, Y . Zhao, H. Liu, and H. Pei, “The infrared and visible image fusion algorithm based on target separation and sparse representation,”Infrared Physics & Technology, vol. 67, pp. 397–407, 2014

    2014

  22. [31]

    Infrared and visible image fusion using visual saliency sparse representation and detail injection model,

    Y . Yang, Y . Zhang, S. Huang, Y . Zuo, and J. Sun, “Infrared and visible image fusion using visual saliency sparse representation and detail injection model,”IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–15, 2020

    2020

  23. [32]

    Joint coupled dictionaries-based visible-infrared image fusion method via texture preservation structure in sparse domain,

    C. Zhang, H. Li, Z. Feng, and S. He, “Joint coupled dictionaries-based visible-infrared image fusion method via texture preservation structure in sparse domain,”Computer Vision and Image Understanding, vol. 235, p. 103781, 2023

    2023

  24. [33]

    Infrared and visible image fusion method based on principal component analysis network and multi-scale morphological gradient,

    S. Li, Y . Zou, G. Wang, and C. Lin, “Infrared and visible image fusion method based on principal component analysis network and multi-scale morphological gradient,”Infrared Physics & Technology, vol. 133, p. 104810, 2023

    2023

  25. [34]

    Novel infrared and visible image fusion method based on independent component analysis,

    Y . Lu, F. Wang, X. Luo, and F. Liu, “Novel infrared and visible image fusion method based on independent component analysis,”Frontiers of Computer Science, vol. 8, pp. 243–254, 2014

    2014

  26. [35]

    Multi-focus image fusion based on non-negative matrix factorization and difference images,

    Y . Zhang, L. Chen, J. Jia, and Z. Zhao, “Multi-focus image fusion based on non-negative matrix factorization and difference images,”Signal Processing, vol. 105, pp. 84–97, 2014

    2014

  27. [36]

    Detail preserved fusion of visible and infrared images using regional saliency extraction and multi-scale image decomposition,

    G. Cui, H. Feng, Z. Xu, Q. Li, and Y . Chen, “Detail preserved fusion of visible and infrared images using regional saliency extraction and multi-scale image decomposition,”Optics Communications, vol. 341, pp. 199–209, 2015

    2015

  28. [37]

    Infrared and visible image fusion based on visual saliency map and weighted least square opti- mization,

    J. Ma, Z. Zhou, B. Wang, and H. Zong, “Infrared and visible image fusion based on visual saliency map and weighted least square opti- mization,”Infrared Physics & Technology, vol. 82, pp. 8–17, 2017. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 17

    2017

  29. [38]

    Infrared and visible image fusion via salient object extraction and low-light region enhancement,

    Y . Liu, L. Dong, and W. Xu, “Infrared and visible image fusion via salient object extraction and low-light region enhancement,”Infrared Physics & Technology, vol. 124, p. 104223, 2022

    2022

  30. [39]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014

    2014

  31. [40]

    Ddcgan: A dual-discriminator conditional generative adversarial network for multi- resolution image fusion,

    J. Ma, H. Xu, J. Jiang, X. Mei, and X.-P. Zhang, “Ddcgan: A dual-discriminator conditional generative adversarial network for multi- resolution image fusion,”IEEE Transactions on Image Processing, vol. 29, pp. 4980–4995, 2020

    2020

  32. [41]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  33. [42]

    A transformer-based framework for automatic covid19 diagnosis in chest cts,

    L. Zhang and Y . Wen, “A transformer-based framework for automatic covid19 diagnosis in chest cts,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 513–518

    2021

  34. [43]

    Toward accurate infrared small target detection via edge-aware gated transformer,

    Y . Zhu, Y . Ma, F. Fan, J. Huang, K. Wu, and G. Wang, “Toward accurate infrared small target detection via edge-aware gated transformer,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 8779–8793, 2024

    2024

  35. [44]

    Ds-transunet: Dual swin transformer u-net for medical image segmentation,

    A. Lin, B. Chen, J. Xu, Z. Zhang, G. Lu, and D. Zhang, “Ds-transunet: Dual swin transformer u-net for medical image segmentation,”IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–15, 2022

    2022

  36. [45]

    Swinfusion: Cross-domain long-range learning for general image fusion via swin transformer,

    J. Ma, L. Tang, F. Fan, J. Huang, X. Mei, and Y . Ma, “Swinfusion: Cross-domain long-range learning for general image fusion via swin transformer,”IEEE/CAA Journal of Automatica Sinica, vol. 9, no. 7, pp. 1200–1217, 2022

    2022

  37. [46]

    Background semantics matter: Cross-task feature exchange network for clustered infrared small target detection with sky-annotated dataset,

    Y . Dai, M. Xiao, Y . Zhu, H. Wang, K. Guo, and J. Yang, “Background semantics matter: Cross-task feature exchange network for clustered infrared small target detection with sky-annotated dataset,” 2024. [Online]. Available: https://arxiv.org/abs/2407.20078

    2024

  38. [47]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  39. [48]

    Dau-net: A regression cell counting method,

    Y . Zhu, S. Tang, Y . Jiang, and R. Kang, “Dau-net: A regression cell counting method,” inISCTT 2021; 6th International Conference on Information Science, Computer Technology and Transportation. VDE, 2021, pp. 1–6

    2021

  40. [49]

    Rdca-net: Residual dense channel attention symmetric network for infrared and visible image fusion,

    Z. Huang, B. Yang, and C. Liu, “Rdca-net: Residual dense channel attention symmetric network for infrared and visible image fusion,” Infrared Physics & Technology, vol. 130, p. 104589, 2023

    2023

  41. [50]

    Ssau-net: A spectral–spatial attention-based u-net for hyperspectral image fusion,

    S. Liu, S. Liu, S. Zhang, B. Li, W. Hu, and Y .-D. Zhang, “Ssau-net: A spectral–spatial attention-based u-net for hyperspectral image fusion,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1– 16, 2022

    2022

  42. [51]

    Attention-aware temporal–spatial graph neural network with multi-sensor information fusion for fault diagnosis,

    Z. Wang, Z. Wu, X. Li, H. Shao, T. Han, and M. Xie, “Attention-aware temporal–spatial graph neural network with multi-sensor information fusion for fault diagnosis,”Knowledge-Based Systems, vol. 278, p. 110891, 2023

    2023

  43. [52]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  44. [53]

    Cbam: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 3–19

    2018

  45. [54]

    Fcanet: Frequency channel attention networks,

    Z. Qin, P. Zhang, F. Wu, and X. Li, “Fcanet: Frequency channel attention networks,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 783–792

    2021

  46. [55]

    Paying more attention to attention: Improving the performance of convolutional neural networks via atten- tion transfer,

    S. Zagoruyko and N. Komodakis, “Paying more attention to attention: Improving the performance of convolutional neural networks via atten- tion transfer,”arXiv preprint arXiv:1612.03928, 2016

    2016 arXiv

  47. [56]

    Image fusion based on pixel significance using cross bilateral filter,

    B. Shreyamsha Kumar, “Image fusion based on pixel significance using cross bilateral filter,”Signal, image and video processing, vol. 9, pp. 1193–1204, 2015

    2015

  48. [57]

    Infrared and visible image fusion using latent low-rank representation,

    H. Li and X.-J. Wu, “Infrared and visible image fusion using latent low-rank representation,”arXiv preprint arXiv:1804.08992, 2018

    2018 arXiv

  49. [58]

    Rethinking the image fusion: A fast unified image fusion network based on proportional maintenance of gradient and intensity,

    H. Zhang, H. Xu, Y . Xiao, X. Guo, and J. Ma, “Rethinking the image fusion: A fast unified image fusion network based on proportional maintenance of gradient and intensity,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 12 797– 12 804

    2020

  50. [59]

    Classification saliency-based rule for visible and infrared image fusion,

    H. Xu, H. Zhang, and J. Ma, “Classification saliency-based rule for visible and infrared image fusion,”IEEE Transactions on Computational Imaging, vol. 7, pp. 824–836, 2021

    2021

  51. [60]

    A multi-scale information integra- tion framework for infrared and visible image fusion,

    G. Yang, J. Li, H. Lei, and X. Gao, “A multi-scale information integra- tion framework for infrared and visible image fusion,”Neurocomputing, vol. 600, p. 128116, 2024

    2024

  52. [61]

    Llvip: A visible-infrared paired dataset for low-light vision,

    X. Jia, C. Zhu, M. Li, W. Tang, and W. Zhou, “Llvip: A visible-infrared paired dataset for low-light vision,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3496–3504

    2021

  53. [62]

    The tno multiband image data collection,

    A. Toet, “The tno multiband image data collection,”Data in brief, vol. 15, pp. 249–251, 2017

    2017

  54. [63]

    Piafusion: A progres- sive infrared and visible image fusion network based on illumination aware,

    L. Tang, J. Yuan, H. Zhang, X. Jiang, and J. Ma, “Piafusion: A progres- sive infrared and visible image fusion network based on illumination aware,”Information Fusion, vol. 83, pp. 79–92, 2022

    2022

  55. [64]

    Fusiondn: A unified densely connected network for image fusion,

    H. Xu, J. Ma, Z. Le, J. Jiang, and X. Guo, “Fusiondn: A unified densely connected network for image fusion,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 12 484–12 491

    2020

  56. [65]

    Multi-spectral sift for scene category recognition,

    M. Brown and S. S ¨usstrunk, “Multi-spectral sift for scene category recognition,” inCVPR 2011. IEEE, 2011, pp. 177–184

    2011

  57. [66]

    Information measure for performance of image fusion,

    G. Qu, D. Zhang, and P. Yan, “Information measure for performance of image fusion,”Electronics letters, vol. 38, no. 7, p. 1, 2002

    2002

  58. [67]

    A nonlinear correlation measure for multivariable data set,

    Q. Wang, Y . Shen, and J. Q. Zhang, “A nonlinear correlation measure for multivariable data set,”Physica D: Nonlinear Phenomena, vol. 200, no. 3-4, pp. 287–295, 2005

    2005

  59. [68]

    Objective image fusion performance measure,

    C. S. Xydeas, V . Petrovicet al., “Objective image fusion performance measure,”Electronics letters, vol. 36, no. 4, pp. 308–309, 2000

    2000

  60. [69]

    Performance assessment of combina- tive pixel-level image fusion based on an absolute feature measurement,

    J. Zhao, R. Laganiere, and Z. Liu, “Performance assessment of combina- tive pixel-level image fusion based on an absolute feature measurement,” Int. J. Innov. Comput. Inf. Control, vol. 3, no. 6, pp. 1433–1447, 2007

    2007

  61. [70]

    A novel similarity based quality metric for image fusion,

    S. Li, R. Hong, and X. Wu, “A novel similarity based quality metric for image fusion,” in2008 International Conference on Audio, Language and Image Processing. IEEE, 2008, pp. 167–172

    2008

  62. [71]

    A new image fusion performance metric based on visual information fidelity,

    Y . Han, Y . Cai, Y . Cao, and X. Xu, “A new image fusion performance metric based on visual information fidelity,”Information fusion, vol. 14, no. 2, pp. 127–135, 2013

    2013

  63. [72]

    Ultralytics yolov5,

    G. Jocher, “Ultralytics yolov5,” 2020. [Online]. Available: https: //github.com/ultralytics/yolov5

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.