pith. machine review for the scientific record. sign in

arxiv: 2604.22886 · v1 · submitted 2026-04-24 · 💻 cs.CV

Breaking Degradation Coupling: A Structural Entropy Guided Decoupled Framework and Benchmark for Infrared Enhancement

Pu Li , Huafeng Li , Yafei Zhang , Yu Liu , Wen Wang This is my paper

Pith reviewed 2026-05-08 12:26 UTC · model grok-4.3

classification 💻 cs.CV
keywords thermal infrared enhancementdegradation decouplingstructural entropyevidential networkresidual modulesimage restorationnighttime benchmark
0
0 comments X

The pith

SEGD decouples compound degradations in thermal infrared images into independent residual modules selected by structural entropy.

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

The paper proposes the Structural Entropy-Guided Decoupled (SEGD) Framework to restore high-quality thermal infrared images from multiple overlapping degradations. It decomposes the problem into separate sub-tasks handled by Degradation-Specific Residual Modules that estimate task-specific corrections, with an evidential network estimating degradation type and strength to control restoration intensity. Multiple possible orderings of these modules create restoration paths whose outputs are aggregated by choosing the most informative features according to a structural entropy measure. This produces decoder-ready representations that retain both structural detail and awareness of the original degradations. The approach is tested against existing methods and supported by a new benchmark of nighttime infrared images.

Core claim

By modeling each degradation type with its own residual module, composing these modules into alternative sequences, and selecting among the resulting feature paths via structural entropy, SEGD yields representations that preserve structural fidelity while remaining aware of the input degradations, enabling finer and more interpretable enhancement than shared-backbone models.

What carries the argument

Degradation-Specific Residual Modules (DRMs) that perform residual estimation for one degradation type at a time, arranged in varying orders to form multiple paths whose outputs are filtered by a structural-entropy criterion after receiving priors from a Degradation-Aware Evidential Network.

Load-bearing premise

Compound degradations can be split into independent sub-processes whose interactions do not need joint modeling, and structural-entropy selection reliably yields features that are both structurally accurate and degradation-aware.

What would settle it

A direct comparison on a dataset of strongly coupled degradations (for example simultaneous low-light and sensor noise where one type alters the statistics of the other) in which SEGD produces lower reconstruction quality or requires more parameters than a single shared-backbone baseline.

Figures

Figures reproduced from arXiv: 2604.22886 by Huafeng Li, Pu Li, Wen Wang, Yafei Zhang, Yu Liu.

Figure 1
Figure 1. Figure 1: Conceptual illustration of the proposed SEGD frame view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed SEGD framework. SEGD integrates degradation perception, residual restoration, and order view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison under triple compound degradations on the HM-TIR and Night-TIR datasets. Due to space constraints, view at source ↗
Figure 4
Figure 4. Figure 4: Quantitative and qualitative comparisons on the AWMM dataset. view at source ↗
Figure 7
Figure 7. Figure 7: Results on different degradation restoration strategies. view at source ↗
Figure 6
Figure 6. Figure 6: The efficacy of each DRM toward specific degradation. view at source ↗
Figure 8
Figure 8. Figure 8: TIR images within Night-TIR benchmark. 8. Network Architectural Details The proposed SEGD framework comprises an encoder, a decoder, DENet, and a set of DRMs. The encoder contains two convolutional layers followed by eight ResBlocks (each residual block comprises two convolutional layers, each fol￾lowed by GroupNorm and GELU). The feature width is fixed at 64 channels. The decoder consists of four Res￾Bloc… view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative comparison of single and double compound degradations on the HM-TIR and Night-TIR datasets. view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative and quantitative comparisons under single-degradation settings on HM-TIR and Night-TIR. view at source ↗
read the original abstract

Thermal infrared image enhancement aims to restore high-quality images from complex compound degradations. Existing all-in-one approaches typically employ a single shared backbone to handle diverse degradations, which causes gradient interference and parameter competition. To address this, we propose a Structural Entropy-Guided Decoupled (SEGD) Framework. Unlike unified modeling paradigms, SEGD decomposes compound degradations into independent sub-processes and models them in a divide-and-conquer manner through Degradation-Specific Residual Modules (DRMs). Each DRM focuses on residual estimation for a specific degradation, enabling task decoupling while remaining jointly trainable, which mitigates parameter contention. A Degradation-Aware Evidential Network further estimates degradation type and intensity, providing priors that adaptively regulate DRM restoration strength. To handle compound cases, DRMs are composed in varying orders to form multiple restoration paths, from which the most informative features are aggregated under a structural-entropy criterion, yielding decoder-ready representations with structural fidelity and degradation awareness. Integrating divide-and-conquer restoration, evidential perception, and entropy-guided adaptation, SEGD achieves fine-grained and interpretable enhancement. We also construct a nighttime TIR benchmark for evaluation under real low-light conditions. Experimental results demonstrate that SEGD surpasses state-of-the-art methods while achieving higher efficiency with fewer parameters.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces the Structural Entropy-Guided Decoupled (SEGD) Framework for thermal infrared (TIR) image enhancement under compound degradations. It decomposes degradations via Degradation-Specific Residual Modules (DRMs) that perform task-specific residual estimation in a jointly trainable manner, employs a Degradation-Aware Evidential Network to estimate degradation type and intensity as adaptive priors, and aggregates features across multiple DRM composition paths using a structural-entropy criterion to produce decoder-ready representations. A new nighttime TIR benchmark is constructed for real low-light evaluation. The central claim is that SEGD outperforms state-of-the-art unified methods in enhancement quality while using fewer parameters and achieving higher efficiency.

Significance. If the reported results and ablations hold, this work offers a substantive contribution to infrared restoration by mitigating gradient interference and parameter competition through explicit decoupling, while adding interpretability via evidential priors and entropy-guided selection. The construction of a dedicated real-world low-light TIR benchmark fills an evaluation gap. Explicit credit is given for the ablation studies on path composition and degradation estimation, which directly probe the core decomposition assumption, and for the parameter-efficient design that is jointly trainable without evident contention.

minor comments (3)
  1. Abstract: the performance claims would be strengthened by including at least one concrete quantitative result (e.g., average PSNR/SSIM gain or parameter count reduction) rather than qualitative statements alone.
  2. Notation: ensure consistent definition and symbol usage for 'structural entropy' and the aggregation operator across the method description and any equations; a short appendix derivation would aid reproducibility.
  3. Benchmark section: specify the exact acquisition conditions, degradation statistics, and train/test split sizes to allow direct comparison with future work.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the careful reading and positive evaluation of our work. The recommendation for minor revision is appreciated, and we are encouraged that the significance of the SEGD framework, the new nighttime TIR benchmark, and the ablation studies on decoupling and path selection have been recognized. No major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The SEGD framework decomposes compound degradations via independently motivated DRMs, an evidential perception network, and structural-entropy path aggregation. These are architectural and algorithmic choices presented as responses to gradient interference in unified models, with no equations or components reducing to self-definition, fitted parameters renamed as predictions, or load-bearing self-citations. The nighttime TIR benchmark is constructed separately for evaluation. Experimental results and ablations provide external validation rather than tautological equivalence, rendering the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

The framework introduces new modules and selection criteria whose validity rests on unstated assumptions about degradation independence and entropy as a fidelity proxy; no explicit free parameters or external axioms are named in the abstract.

invented entities (2)
  • Degradation-Specific Residual Modules (DRMs) no independent evidence
    purpose: Focus on residual estimation for one specific degradation type to enable task decoupling
    Introduced to mitigate parameter competition in unified backbones
  • Degradation-Aware Evidential Network no independent evidence
    purpose: Estimate degradation type and intensity to provide adaptive priors
    Regulates restoration strength of the DRMs

pith-pipeline@v0.9.0 · 5534 in / 1278 out tokens · 50151 ms · 2026-05-08T12:26:03.548159+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 4 canonical work pages

  1. [1]

    Multimodal prompt perceiver: Empower adap- tiveness generalizability and fidelity for all-in-one image restoration

    Yuang Ai, Huaibo Huang, Xiaoqiang Zhou, Jiexiang Wang, and Ran He. Multimodal prompt perceiver: Empower adap- tiveness generalizability and fidelity for all-in-one image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25432– 25444, 2024. 2

    2024

  2. [2]

    Why thermal im- ages are blurry.Optics Express, 32(3):3852–3865, 2024

    Fanglin Bao, Shubhankar Jape, Andrew Schramka, Junjie Wang, Tim E McGraw, and Zubin Jacob. Why thermal im- ages are blurry.Optics Express, 32(3):3852–3865, 2024. 2

    2024

  3. [3]

    Exploring video denoising in thermal infrared imaging: Physics-inspired noise generator, dataset, and model.IEEE Transactions on Image Processing, 33:3839–3854, 2024

    Lijing Cai, Xiangyu Dong, Kailai Zhou, and Xun Cao. Exploring video denoising in thermal infrared imaging: Physics-inspired noise generator, dataset, and model.IEEE Transactions on Image Processing, 33:3839–3854, 2024. 2

    2024

  4. [4]

    Remote sensing image stripe noise removal: From image decompo- sition perspective.IEEE Transactions on Geoscience and Remote Sensing, 54(12):7018–7031, 2016

    Yi Chang, Luxin Yan, Tao Wu, and Sheng Zhong. Remote sensing image stripe noise removal: From image decompo- sition perspective.IEEE Transactions on Geoscience and Remote Sensing, 54(12):7018–7031, 2016. 2, 6

    2016

  5. [5]

    Score priors guided deep variational inference for unsupervised real-world single im- age denoising

    Jun Cheng, Tao Liu, and Shan Tan. Score priors guided deep variational inference for unsupervised real-world single im- age denoising. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 12937–12948,

  6. [6]

    Rethinking coarse-to-fine approach in sin- gle image deblurring

    Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in sin- gle image deblurring. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4641–4650,

  7. [7]

    In- structir: High-quality image restoration following human in- structions

    Marcos V Conde, Gregor Geigle, and Radu Timofte. In- structir: High-quality image restoration following human in- structions. InProceedings of the European Conference on Computer Vision, pages 1–21. Springer, 2024. 2

    2024

  8. [8]

    Darkir: Robust low-light image restoration

    Daniel Feijoo, Juan C Benito, Alvaro Garcia, and Marcos V Conde. Darkir: Robust low-light image restoration. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10879–10889, 2025. 3

    2025

  9. [9]

    Onerestore: A universal restoration framework for composite degradation

    Yu Guo, Yuan Gao, Yuxu Lu, Huilin Zhu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. InProceedings of the European Conference on Computer Vision, pages 255–272. Springer, 2024. 2

    2024

  10. [10]

    Nonuniformity correc- tion of infrared image sequences using the constant-statistics constraint.IEEE Transactions on Image Processing, 8(8): 1148–1151, 1999

    John G Harris and Yu-Ming Chiang. Nonuniformity correc- tion of infrared image sequences using the constant-statistics constraint.IEEE Transactions on Image Processing, 8(8): 1148–1151, 1999. 2

    1999

  11. [11]

    Single-image-based nonuniformity correction of uncooled long-wave infrared de- tectors: A deep-learning approach.Applied optics, 57(18): D155–D164, 2018

    Zewei He, Yanpeng Cao, Yafei Dong, Jiangxin Yang, Yan- long Cao, and Christel-L ¨oic Tisse. Single-image-based nonuniformity correction of uncooled long-wave infrared de- tectors: A deep-learning approach.Applied optics, 57(18): D155–D164, 2018. 1

    2018

  12. [12]

    Uni- versal image restoration pre-training via degradation classi- fication

    JiaKui Hu, Lujia Jin, Zhengjian Yao, and Yanye Lu. Uni- versal image restoration pre-training via degradation classi- fication. InProceedings of the International Conference on Learning Representations, pages 1–16, 2025. 2, 4, 5

    2025

  13. [13]

    Infrared thermal image denoising with symmetric multi-scale sampling network.Infrared Physics & Technol- ogy, 134:104909, 2023

    Xinrui Hu, Shaojuan Luo, Chunhua He, Wenhao Wu, and Heng Wu. Infrared thermal image denoising with symmetric multi-scale sampling network.Infrared Physics & Technol- ogy, 134:104909, 2023. 2

    2023

  14. [14]

    Thermal wave image deblurring based on depth residual network.Infrared Physics & Technology, 117: 103847, 2021

    Haijun Jiang, Fei Chen, Xining Liu, Jesse Chen, Kai Zhang, and Li Chen. Thermal wave image deblurring based on depth residual network.Infrared Physics & Technology, 117: 103847, 2021. 2

    2021

  15. [15]

    An in- frared thermal image denoising method focusing on noise feature learning.Optics & Laser Technology, 184:112475,

    Nanhe Jiang, Yucun Zhang, Qun Li, and Fang Yan. An in- frared thermal image denoising method focusing on noise feature learning.Optics & Laser Technology, 184:112475,

  16. [16]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148–5157, 2021. 6

    2021

  17. [17]

    Efficient frequency domain-based trans- formers for high-quality image deblurring

    Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based trans- formers for high-quality image deblurring. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 5886–5895, 2023. 4

    2023

  18. [18]

    Efficient visual state space model for image deblurring

    Lingshun Kong, Jiangxin Dong, Jinhui Tang, Ming-Hsuan Yang, and Jinshan Pan. Efficient visual state space model for image deblurring. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 12710–12719, 2025. 4

    2025

  19. [19]

    Efficient diffusion as low light enhancer

    Guanzhou Lan, Qianli Ma, Yuqi Yang, Zhigang Wang, Dong Wang, Xuelong Li, and Bin Zhao. Efficient diffusion as low light enhancer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21277– 21286, 2025. 3

    2025

  20. [20]

    Ap- bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network

    Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. Ap- bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17725–17734, 2022. 4

    2022

  21. [21]

    Structural information and dynamical complexity of networks.IEEE Transactions on Information Theory, 62(6):3290–3339, 2016

    Angsheng Li and Yicheng Pan. Structural information and dynamical complexity of networks.IEEE Transactions on Information Theory, 62(6):3290–3339, 2016. 2, 5, 1

    2016

  22. [22]

    All-weather multi-modality image fu- sion: Unified framework and 100k benchmark.Information Fusion, page 104130, 2026

    Xilai Li, Wuyang Liu, Xiaosong Li, Fuqiang Zhou, Huafeng Li, and Feiping Nie. All-weather multi-modality image fu- sion: Unified framework and 100k benchmark.Information Fusion, page 104130, 2026. 6

    2026

  23. [23]

    arXiv preprint arXiv:2312.05038 (2023) MMFE-IR 17

    Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration.arXiv preprint arXiv:2312.05038, 2023. 2

    work page arXiv 2023

  24. [24]

    Enhancing infrared vision: Progressive prompt fusion network and benchmark

    Jinyuan Liu, Zihang Chen, Zhu Liu, Zhiying Jiang, Long Ma, Xin Fan, and Risheng Liu. Enhancing infrared vision: Progressive prompt fusion network and benchmark. InAd- vances in Neural Information Processing Systems, pages 1–

  25. [25]

    1, 3, 6, 5

    Curran Associates, Inc., 2025. 1, 3, 6, 5

    2025

  26. [26]

    Li Liu, Luping Xu, and Houzhang Fang. Simultaneous in- tensity bias estimation and stripe noise removal in infrared images using the global and local sparsity constraints.IEEE Transactions on Geoscience and Remote Sensing, 58(3): 1777–1789, 2019. 2

    2019

  27. [27]

    Vl-ur: Vision-language-guided universal restoration of images de- graded by adverse weather conditions

    Ziyan Liu, Yuxu Lu, Hushan Yu, and Dong Yang. Vl-ur: Vision-language-guided universal restoration of images de- graded by adverse weather conditions. In2025 IEEE Inter- national Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2025. 2

    2025

  28. [28]

    Deal: Data-efficient adversarial learn- ing for high-quality infrared imaging

    Zhu Liu, Zijun Wang, Jinyuan Liu, Fanqi Meng, Long Ma, and Risheng Liu. Deal: Data-efficient adversarial learn- ing for high-quality infrared imaging. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28198–28207, 2025. 1, 2, 3

    2025

  29. [29]

    Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B

    Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B. Sch¨on. Controlling vision-language models for multi-task image restoration. InProceedings of the International Conference on Learning Representations, pages 1–13, 2024. 2, 6

    2024

  30. [30]

    Prores: Exploring degradation-aware visual prompt for universal image restora- tion.arXiv preprint arXiv:2306.13653, 2023

    Jiaqi Ma, Tianheng Cheng, Guoli Wang, Qian Zhang, Xinggang Wang, and Lefei Zhang. Prores: Exploring degradation-aware visual prompt for universal image restora- tion.arXiv preprint arXiv:2306.13653, 2023. 2

    work page arXiv 2023

  31. [31]

    Learning with self- calibrator for fast and robust low-light image enhancement

    Long Ma, Tengyu Ma, Chengpei Xu, Jinyuan Liu, Xin Fan, Zhongxuan Luo, and Risheng Liu. Learning with self- calibrator for fast and robust low-light image enhancement. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2025. 3

    2025

  32. [32]

    Allrestorer: All-in-one transformer for image restoration under composite degradations,

    Jiawei Mao, Yu Yang, Xuesong Yin, Ling Shao, and Hao Tang. Allrestorer: All-in-one transformer for image restoration under composite degradations.arXiv preprint arXiv:2411.10708, 2024. 2

    work page arXiv 2024

  33. [33]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Mak- ing a “completely blind” image quality analyzer.IEEE Sig- nal Processing Letters, 20(3):209–212, 2012. 6

    2012

  34. [34]

    Stripe and ring artifact removal with combined wavelet—fourier filtering.Optics express, 17(10):8567– 8591, 2009

    Beat M ¨unch, Pavel Trtik, Federica Marone, and Marco Stam- panoni. Stripe and ring artifact removal with combined wavelet—fourier filtering.Optics express, 17(10):8567– 8591, 2009. 2, 6

    2009

  35. [35]

    An infrared image enhance- ment method via content and detail two-stream deep con- volutional neural network.Infrared Physics & Technology, 132:104761, 2023

    Zhongxiang Pang, Guihua Liu, Guosheng Li, Jian Gong, Chunmei Chen, and Chao Yao. An infrared image enhance- ment method via content and detail two-stream deep con- volutional neural network.Infrared Physics & Technology, 132:104761, 2023. 1, 3, 6

    2023

  36. [36]

    In- frared image dynamic range compression based on adaptive contrast adjustment and structure preservation.IEEE Trans- actions on Geoscience and Remote Sensing, 2024

    Jinyi Qiu, Zhan Wang, Yuanfei Huang, and Hua Huang. In- frared image dynamic range compression based on adaptive contrast adjustment and structure preservation.IEEE Trans- actions on Geoscience and Remote Sensing, 2024. 2

    2024

  37. [37]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InProceedings of the International Conference on Machine Learning, pages 8748–8763. PmLR, 2021. 2

    2021

  38. [38]

    Manifold-aware representation learning for degradation-agnostic image restoration.arXiv preprint arXiv:2505.18679, 2025

    Bin Ren, Yawei Li, Xu Zheng, Yuqian Fu, Danda Pani Paudel, Ming-Hsuan Yang, Luc Van Gool, and Nicu Sebe. Manifold-aware representation learning for degradation-agnostic image restoration.arXiv preprint arXiv:2505.18679, 2025. 2

    work page arXiv 2025

  39. [39]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 2

    2022

  40. [40]

    Eviden- tial deep learning to quantify classification uncertainty

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Eviden- tial deep learning to quantify classification uncertainty. In Advances in Neural Information Processing Systems, pages 1–11, 2018. 2, 3, 4, 1

    2018

  41. [41]

    Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018

    Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018. 6

    2018

  42. [42]

    Stripformer: Strip transformer for fast image deblurring

    Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. InProceedings of the European Confer- ence on Computer Vision, pages 146–162. Springer, 2022. 4

    2022

  43. [43]

    Fast motion-deblurring of ir images.IEEE Signal Processing Letters, 29:459–463, 2022

    Nisha Varghese, Mahesh Mohan MR, and AN Rajagopalan. Fast motion-deblurring of ir images.IEEE Signal Processing Letters, 29:459–463, 2022. 2

    2022

  44. [44]

    A real-time contrast enhancement algorithm for in- frared images based on plateau histogram.Infrared Physics & Technology, 48(1):77–82, 2006

    Bing-jian Wang, Shang-Qian Liu, Qing Li, and Hui-Xin Zhou. A real-time contrast enhancement algorithm for in- frared images based on plateau histogram.Infrared Physics & Technology, 48(1):77–82, 2006. 2

    2006

  45. [45]

    Target attention deep neural network for infrared image enhancement.Infrared Physics & Technology, 115:103690, 2021

    Dong Wang, Rui Lai, and Juntao Guan. Target attention deep neural network for infrared image enhancement.Infrared Physics & Technology, 115:103690, 2021. 2

    2021

  46. [46]

    Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method

    Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 2654–2662, 2023. 3

    2023

  47. [47]

    Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process- ing, 13(4):600–612, 2004. 4, 6

    2004

  48. [48]

    Thermal-aware low-light image enhancement: A real-world benchmark and a new light-weight model

    Zhen Wang, Yaozu Wu, Dongyuan Li, Shiyin Tan, and Zhishuai Yin. Thermal-aware low-light image enhancement: A real-world benchmark and a new light-weight model. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 8223–8231, 2025. 2

    2025

  49. [49]

    Community detection in large-scale com- plex networks via structural entropy game

    Yantuan Xian, Pu Li, Hao Peng, Zhengtao Yu, Yan Xiang, and Philip S Yu. Community detection in large-scale com- plex networks via structural entropy game. InProceedings of the 2025 ACM Web Conference (WWW), pages 3930–3941. ACM, 2025. 2, 5, 1, 6

    2025

  50. [50]

    Hctirde- blur: A hybrid convolution-transformer network for single infrared image deblurring.Infrared Physics & Technology, 131:104640, 2023

    Shi Yi, Li Li, Xi Liu, Junjie Li, and Ling Chen. Hctirde- blur: A hybrid convolution-transformer network for single infrared image deblurring.Infrared Physics & Technology, 131:104640, 2023. 2

    2023

  51. [51]

    Com- plexity experts are task-discriminative learners for any im- age restoration

    Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yuedong Tan, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Com- plexity experts are task-discriminative learners for any im- age restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753– 12763, 2025. 2, 4, 5

    2025

  52. [52]

    Cycleisp: Real image restoration via improved data synthesis

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Cycleisp: Real image restoration via improved data synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2696– 2705, 2020. 4

    2020

  53. [53]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728– 5739, 2022. 4, 2

    2022

  54. [54]

    A combined stripe noise removal and deblurring recovering method for thermal infrared remote sensing images.IEEE Transactions on Geoscience and Re- mote Sensing, 60:1–14, 2022

    Jingwen Zhang, Xiaoxuan Zhou, Liyuan Li, Tingliang Hu, and Chen Fansheng. A combined stripe noise removal and deblurring recovering method for thermal infrared remote sensing images.IEEE Transactions on Geoscience and Re- mote Sensing, 60:1–14, 2022. 2

    2022

  55. [55]

    Selective hourglass mapping for universal image restoration based on diffusion model

    Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25445– 25455, 2024. 2, 6

    2024

  56. [56]

    Thermal infrared spectrometer on-orbit de- focus assessment based on blind image blur kernel estima- tion.Infrared Physics & Technology, 130:104538, 2023

    Xiaoxuan Zhou, Jingwen Zhang, Mao Li, Xiaofeng Su, and Fansheng Chen. Thermal infrared spectrometer on-orbit de- focus assessment based on blind image blur kernel estima- tion.Infrared Physics & Technology, 130:104538, 2023. 2

    2023

  57. [57]

    Iterative de- noiser and noise estimator for self-supervised image denois- ing

    Yunhao Zou, Chenggang Yan, and Ying Fu. Iterative de- noiser and noise estimator for self-supervised image denois- ing. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 13265–13274, 2023. 4 Breaking Degradation Coupling: A Structural Entropy–Guided Decoupled Framework and Benchmark for Infrared Enhancement Supplementary Material

    2023

  58. [58]

    Preliminaries In this section, we summarize the concepts and definitions related to the background of our work, including thermal in- frared (TIR) image enhancement, Evidential deep learning, and Structural entropy. 6.1. TIR Enhancement TIR degradations primarily manifest as low contrast, blur, and noise. These effects arise from (i) insufficient tar- get...

  59. [59]

    For the root nodeλofT,T λ =V

    Each nodexinTis associated with a setT x ⊆ V. For the root nodeλofT,T λ =V. Any leaf nodeqinTis associated with a single node inG, i.e.,T q ={v},v∈ V

  60. [60]

    , bk, thenT b1 ,

    For each nodeainT, denote all its children as b1, . . . , bk, thenT b1 , . . . , Tbk is a partition ofT a

  61. [61]

    Let h(λ) = 0andh(¯a) =h(a) + 1, where¯ais the parent of a

    For each nodeainT, denote its height ash(a). Let h(λ) = 0andh(¯a) =h(a) + 1, where¯ais the parent of a. The height ofT,h(T) = max a∈T h(a). The SE of graphGon coding treeTis defined as: HT (G) =− X a∈T,a̸=λ ga vol(λ) log vol(a) vol(¯a),(20) whereg a is the summation of the degrees of the cut edges ofT a (i.e., the weight sum of edges with exactly one end-...

  62. [62]

    To this end, we introduce a challenging nighttime TIR enhancement benchmark, Night- TIR

    Additional Details of Night-TIR Benchmark Considering that nighttime scenes exhibit smaller tar- get–background temperature differences and weaker radia- tive signals, thermal infrared (TIR) imagery therefore tends to have reduced contrast. To this end, we introduce a challenging nighttime TIR enhancement benchmark, Night- TIR. As shown in Figure 8, Night...

  63. [63]

    The encoder contains two convolutional layers followed by eight ResBlocks (each residual block comprises two convolutional layers, each fol- lowed by GroupNorm and GELU)

    Network Architectural Details The proposed SEGD framework comprises an encoder, a decoder, DENet, and a set of DRMs. The encoder contains two convolutional layers followed by eight ResBlocks (each residual block comprises two convolutional layers, each fol- lowed by GroupNorm and GELU). The feature width is fixed at 64 channels. The decoder consists of fo...

  64. [64]

    Figure 9

    More Results on HM-TIR and Night-TIR To thoroughly evaluate the proposed SEGD, this section presents three additional studies: (i) qualitative compar- isons with competing methods under single- and double- degradation settings; (ii) comparisons betweenSEGDand state-of-the-art single-degradation methods in the single- degradation regime; and (iii) evaluati...

  65. [65]

    As summarized in Table 5, SEGD has the fewest parameters among all methods and achieves the second-fastest inference

    Complexity Comparison For a fair comparison, we compute and report the number of learnable parameters, inference time, and floating-point operations (FLOPs) for SEGD and all competing meth- ods on single-degradation HM-TIR inputs at a resolution of640×512, excluding the non–deep-learning baselines WFAF and LRSID. As summarized in Table 5, SEGD has the few...

  66. [66]

    We first examine training strategies for DENet: the DAH and DEH are trained either independently—each Table 4

    Additional Ablation Studies All experiments in this section are conducted on the HM- TIR dataset. We first examine training strategies for DENet: the DAH and DEH are trained either independently—each Table 4. Quantitative comparison with additional visible all-in-one methods on the HM-TIR and Night-TIR datasets. The best and second- best performances for ...

  67. [67]

    Limitations and Future Work We follow the degradation synthesis strategy designed in PPFN [24], where TIR degradations are categorized into low contrast, blur, and noise, and training samples are gen- erated accordingly. However, as noted in PPFN, obtaining strictly paired degraded–clean TIR images is inherently dif- ficult, and any degradation pipeline c...