Pith. sign in

REVIEW 3 major objections 6 minor 47 references

A bio-inspired transformer with a spiking noise gate and one-way high-to-low frequency guidance surpasses prior HDR reconstruction methods on perceptual benchmarks, using a fraction of the parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:52 UTC pith:5PSY5KDB

load-bearing objection Plausible bio-inspired HDR architecture, but the SOTA claim is under-supported by the evidence as presented. the 3 major comments →

arxiv 2607.17456 v1 pith:5PSY5KDB submitted 2026-07-20 cs.CV

Bio-SFT: Asymmetric Cortical Guidance and Retinal Adaptation for Robust HDR Reconstruction

Tingyu Cheng , Ting Zhang , Chongyi Li , Zhaoqing Pan , Tiesong Zhao This is my paper
classification cs.CV
keywords HDR reconstructionsingle-image HDRbio-inspired visionspiking neural networkstransformerNaka-Rushton adaptationasymmetric guidanceHDRTV1K
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that single-image HDR reconstruction is best served by mimicking the retina and visual cortex: a learnable Naka–Rushton front end compresses dynamic range, an explicit Parvo (high-frequency) to Magno (low-frequency) one-way guidance prevents dark-region quantization noise from contaminating structure, and an integrate-and-fire spiking gate acts as a hard dead zone that blocks weak noise. On the HDRTV1K benchmark, this design reports a new state of the art (PSNR 38.53 dB, HDR-VDP-3 8.805, ΔE_ITP 7.93) with only 0.48M parameters, and a 0.14M variant that stays competitive. The case matters because it suggests asymmetric, thresholded information flow is a more efficient inductive bias than symmetric attention for pushing perceptual quality in low-level vision.

Core claim

The central claim is that the bottleneck in SDR-to-HDR is not capacity but information routing: symmetric or bidirectional frequency guidance lets low-frequency quantization noise leak into high-frequency structure, producing ghosting and color shifts. Bio-SFT enforces a unidirectional Parvo-to-Magno flow, where high-frequency features are first purified by an event-driven SNN hard gate (IF neurons with a sparsity prior) and then modulate the low-frequency stream via a channel-wise attention map. The paper reports that this mechanism, combined with a learnable Naka–Rushton adaptation and ON/OFF Difference-of-Gaussians filtering, yields the best perceptual metrics on HDRTV1K (PSNR 38.53 dB, H

What carries the argument

The Parvo–Magno Guidance Block (PMGB), which decouples features into high-frequency (Parvo) and low-frequency (Magno) streams with a content-adaptive filter, then applies the Asymmetric Parvo-to-Magno guidance (H→L): the purified high-frequency stream generates a channel-wise attention map that modulates the low-frequency stream via F*_L = F^_L + β·(σ(M(AvgPool(F^_H))) ⊙ F^_L). The Event-Driven SNN Hard Gating module, using IF dynamics V(t+1)=βV(t)+F(t)−S(t)V_th with a Heaviside spike and a firing-rate sparsity loss, creates a sub-threshold noise dead zone that blocks low-amplitude quantization artifacts before they reach the guidance path.

Load-bearing premise

The load-bearing premise is that the HDRTV1K comparison is run-to-run stable and fair: baselines are said to share the same training protocol 'whenever possible,' but the reported gains (0.17 dB PSNR, 0.054 HDR-VDP-3) have no error bars or significance tests, so if baseline numbers were taken from their papers or the 117-frame test set is noisy, the state-of-the-art claim could invert.

What would settle it

Retrain the top baselines and Bio-SFT under identical random crops, seeds, and schedules on HDRTV1K with at least five seeds; if the PSNR and HDR-VDP-3 gaps shrink below one standard deviation or the order flips, the SOTA claim fails. Separately, ablate the SNN gate with a fixed hard-threshold or soft-shrinkage operator; matching performance would falsify the claim that spiking dynamics are essential.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, one-way high-to-low guidance is a better inductive bias than symmetric or bidirectional fusion for restoration tasks where low-frequency noise corrupts structure.
  • The IF spiking gate acts as a trainable, non-parametric denoiser, offering a way to suppress noise without explicit noise modeling.
  • The 0.14M-parameter variant keeps competitive quality, indicating bio-inspired gating can substitute for raw capacity, with implications for 4K real-time deployment.
  • The learnable Naka–Rushton front end outperforms fixed gamma or linear preprocessing, suggesting adaptive photoreceptor-style compression is useful before deep feature extraction.
  • Cross-dataset results on SI-HDR indicate the design generalizes beyond PQ-encoded broadcast content to traditional inverse tone mapping.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported margins over the best baseline (0.17 dB PSNR, 0.054 HDR-VDP-3) come without error bars or multiple seeds, so a multi-seed rerun could place the gains within noise; the asymmetry principle might survive while the exact SOTA claim would need qualification.
  • A natural test is to replace the SNN gate with a simpler learnable hard threshold (e.g., soft shrinkage) — if that matches Bio-SFT, the essential ingredient is thresholding, not spiking dynamics per se.
  • The same H→L asymmetric guidance could be applied to video HDR with a temporal dimension, where spatio-temporal ghosting is an open problem the paper explicitly defers.
  • Because the gate's dead zone is a learned threshold, its optimal setting likely depends on sensor noise characteristics; a domain-adaptation extension that predicts thresholds from noise statistics would be a logical next step.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes Bio-SFT, a transformer architecture for single-image SDR-to-HDR reconstruction, combining a learnable Naka–Rushton retinal-adaptation frontend, a Parvo/Magno frequency split with asymmetric Parvo-to-Magno guidance, and an event-driven SNN hard-gating module trained with a sparsity prior. Two variants are claimed: Bio-SFT-Lite (0.14M parameters) and the standard Bio-SFT (0.48M). On HDRTV1K the standard variant is reported to outperform HDRTVNet++ (PSNR 38.53 vs 38.36, ΔE_ITP 7.93 vs 8.28, HDR-VDP-3 8.805 vs 8.751), while the Lite variant is said to remain competitive at 0.14M. Results on SI-HDR and a 4K efficiency comparison are also reported.

Significance. If the headline SOTA claim were securely established, the paper would make a useful, if incremental, contribution: a lightweight, end-to-end trainable HDR architecture with bio-inspired components. The paper contains strengths: the ablation tables are internally consistent, the descriptions of datasets and preprocessing are unusually specific, and the cross-dataset SI-HDR comparison is a welcome check on generalization. However, the evidence for the central claim is currently inadequate: the comparison against the strongest baseline is not part of the stated protocol, no variance or significance information is given, and the SNN module as written does not implement the noise-suppression mechanism described in the motivation. These issues are load-bearing because the architecture's value proposition rests on the claimed SOTA margin and on the mechanistic role of hard gating. The paper should be revised before publication.

major comments (3)
  1. [§IV-A, §IV-B, Table I] HDRTVNet++ is the closest competitor and the one used to support the SOTA claim, but it is not listed among the 'ten restoration pipelines' in §IV-A. The training-protocol statement says 'the same training protocol is used for all compared models whenever possible,' leaving open whether the HDRTVNet++ numbers were produced by retraining or copied from the original paper. With 117 test frames and no error bars, repeated seeds, or significance tests, margins of 0.17 dB PSNR, 0.054 HDR-VDP-3, and 0.35 ΔE_ITP are within plausible run-to-run variation. Please retrain all baselines under the identical protocol with multiple seeds, or report variances and statistical tests.
  2. [§III-D, Eq. (15), Fig. 5] The 'hard gating' is implemented as Fout = F + η·Conv1×1(S). If S = 0 in the claimed noise dead zone, then Fout = F, so low-amplitude quantization noise is not 'permanently blocked' or removed from the stream; it is merely not additively enhanced. The dead-zone story in Fig. 5 is therefore not realized by the equation as written, unless an additional masking or multiplicative operation is intended but not presented. Please state the actual gating operation and show how sub-threshold noise is removed, or revise the mechanism description and the title 'hard gating.'
  3. [Tables I, III, IV, V, VI] The paper introduces two configurations, but the identity of the evaluated model shifts. Table I reports the SOTA Bio-SFT with 0.48M parameters, while Table III labels 'Bio-SFT (Ours)' with 0.14M and reproduces the Bio-SFT-Lite metric (38.22 dB). All ablations (Tables IV–VI) report PSNR 38.22 or lower, i.e., they are run on the Lite variant, not on the 0.48M configuration that produces the headline result. Consequently, the contributions of the retinal frontend, SNN gating, and asymmetric guidance to the claimed SOTA performance are not directly tested. Please run ablations on the 0.48M configuration, or explicitly and consistently report which variant supports each claim.
minor comments (6)
  1. [Abstract and §I] The abstract and contributions repeatedly state that Bio-SFT has 0.14M parameters, while Table I presents the standard Bio-SFT with 0.48M and Bio-SFT-Lite with 0.14M. Please harmonize the terminology and parameter counts throughout.
  2. [Table I] The HDRNet row reports PSNR 35.73 but PU-PSNR 27.53, and Deep SR-ITM reports PSNR 37.10 but PU-PSNR 29.98. These values do not follow the pattern of the other rows and look like transcription errors; please verify.
  3. [§IV-A] The implementation paragraph refers to 'the progressive strategy described in Sec. III,' but no progressive training strategy appears in Section III. Please add the description or remove the reference.
  4. [§V and Fig. 5] Fig. 5 describes the mechanism as enabling 'zero-shot denoising,' but the threshold Vth is learned and the sparsity loss λspar is tuned. The paper itself later notes that threshold margins may need tuning for unseen sensors. Please remove 'zero-shot' or provide a dedicated zero-shot experiment.
  5. [§IV-D, Table III] The text says Bio-SFT 'requires exactly 10% of the parameter budget' of HDRTVNet, but 0.14M / 1.41M ≈ 9.9%. Please use 'about' or give the exact ratio.
  6. [Table II] The SI-HDR table labels the model 'Bio-SFT (Ours)' with 0.14M parameters. Please specify whether this is Bio-SFT-Lite or the standard Bio-SFT with a reduced patch size, so the reader can interpret the cross-dataset result.

Circularity Check

0 steps flagged

No load-bearing circularity: the SOTA claim is an empirical benchmark comparison; only a peripheral self-citation is present.

full rationale

The paper's central derivation is not circular. The three components are defined by explicit formulas (Naka-Rushton adaptation in Eqs. 2-4, IF neuron dynamics in Eqs. 12-14, asymmetric Parvo-to-Magno guidance in Eq. 10) whose parameters are learned end-to-end against a held-out test split. The SOTA claim is an empirical comparison on HDRTV1K, not a consequence of those definitions. The only self-citation is Ref. [27] (Gao, Yang, Li, Li), used in Section II-B to note prior bio-prior work on color constancy; it is peripheral and does not justify any design choice or result. The paper candidly states limitations (e.g., Sec. V: 'the explicit thresholding in our SNN gating introduces hyperparameter sensitivity'; Sec. IV-E: 'The visualization does not prove strict denoising behavior in a formal sense'), and it does not import any uniqueness theorem or ansatz via self-citation. The reviewer concerns about the HDRTVNet++ comparison - no error bars, 'the same training protocol is used for all compared models whenever possible,' and Table III using the 0.14M variant for the efficiency row - are statistical-reporting and reproducibility risks, not circularity. Thus no load-bearing circular step exists; the score of 2 reflects only the one minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The ledger shows a modest number of hand-set hyperparameters (two loss weights, unspecified sparsity target) and several domain assumptions about benchmark validity and metric fidelity. No new physical entities are introduced; the 'retina', 'Parvo/Magno', and 'SNN' terms denote computational modules, not claims about biological mechanism. The two loss weights are the only obvious free parameters tuned to make the objective work.

free parameters (5)
  • lambda_grad (gradient loss weight) = 0.05
    Empirically determined loss weighting stated in Sec. III-E; influences structural preservation and hence HDR-VDP-3.
  • lambda_spar (sparsity loss weight) = 0.01
    Empirically determined loss weight for firing-rate sparsity in Sec. III-E; directly controls the SNN dead zone behavior.
  • SNN threshold Vth = learnable
    Spiking threshold in Eqs. 12-13; the dead-zone property depends on its value, learned with a surrogate gradient.
  • sparsity target r_target = not specified
    Target average firing rate in Lspar (Sec. III-D); the paper does not state its value, though the dead-zone effect depends on it.
  • Naka-Rushton parameters n', sigma', and gate parameters k, tau = learned parameter maps / k=10, tau=0.1 default
    Spatially adaptive parameters in the retinal frontend (Sec. III-B); these are learned from data and shape dynamic range compression. The defaults k=10, tau=0.1 are chosen by hand.
axioms (4)
  • domain assumption HDRTV1K and SI-HDR test pairs reliably reflect SDR-to-HDR reconstruction quality.
    Sec. IV-A treats the 4K frames as statistically independent test samples; the central SOTA claim rests on the representativeness and low noise of this benchmark.
  • domain assumption PU21 / HDR-VDP-3 / DeltaE_ITP metrics are faithful proxies for perceived HDR quality.
    The paper's claim of 'competitive perceptual quality' rests on these metrics being meaningful for HDR; Sec. IV-A.
  • ad hoc to paper IF-neuron sub-threshold integration creates a noise dead zone that suppresses quantization noise in dark regions.
    Sec. III-D asserts the mechanism; it is the causal story for the SNN module's benefit, but the paper states the visualization 'does not prove strict denoising behavior' (Sec. IV-E).
  • domain assumption The PQ/Rec.2100 pipeline can be inverted without explicit CRF modeling (bypass through EOTF).
    Sec. IV-A says CRF correction is bypassed because the data follow BT.2100; the approach presumes this simplification does not handicap reconstruction.

pith-pipeline@v1.3.0-alltime-deepseek · 15487 in / 14055 out tokens · 119982 ms · 2026-08-01T17:52:12.828439+00:00 · methodology

0 comments
read the original abstract

Recovering high dynamic range (HDR) radiance from a single standard dynamic range (SDR) image is highly ill-posed. Extreme luminance variation and severe quantization in dark regions make accurate reconstruction challenging, often leading to visual artifacts and color distortions. To address this problem, we propose Bio-SFT, a bio-inspired spiking frequency transformer for single-image HDR reconstruction. Bio-SFT incorporates three biologically motivated components. First, a learnable Naka--Rushton retinal adaptation frontend stabilizes the input under complex lighting conditions. Second, an explicit Parvo--Magno split introduces asymmetric Parvo-to-Magno guidance, allowing high-frequency structural cues to modulate low-frequency reconstruction. Third, an event-driven SNN hard gating module applies all-or-none spiking to suppress dark-region noise while preserving structural details. The module is trained with a sparsity prior to encourage efficient feature utilization. Built for end-to-end training within a transformer backbone, these lightweight components provide strong parameter efficiency. Experiments on HDRTV1K show that Bio-SFT achieves competitive perceptual quality and consistently improves HDR-VDP-3 and $\Delta E_{ITP}$ while reducing artifact propagation in symmetric guidance pipelines.

Figures

Figures reproduced from arXiv: 2607.17456 by Chongyi Li, Tiesong Zhao, Tingyu Cheng, Ting Zhang, Zhaoqing Pan.

Figure 1
Figure 1. Figure 1: Overview of our bio-inspired HDR framework. (a) Conventional [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Biological motivation for decoupling signals in HDR imaging. The [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overall architecture of the Bio-SFT framework. The pipeline aligns with the biological visual system through three sequential stages: (1) a [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the Physiological Retinal Adaptation Front-end. The mod [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Mechanistic comparison of gating strategies. (a) Traditional continuous soft gating (e.g., ReLU) indiscriminately leaks low-amplitude quantization [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: SDR-to-HDRTV pixel luminance mapping. The predicted curve follows [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison against state-of-the-art methods. Bio-SFT produces visually consistent colors and preserves fine structural details under [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparison on an extremely illuminated scene. Bio-SFT reduces [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Pixel-wise absolute error maps visualized with a Jet colormap. Blue [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Scanline intensity profiles over saturated and dark regions. Bio-SFT follows the ground-truth intensity trend more closely. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual comparison on the SI-HDR dataset (glasses scene). Bio-SFT uniquely avoids overexposure and color shifts, maximally preserving detailed lens [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Channel-wise spatial activation maps in challenging dark scenes. Warm [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗

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

47 extracted references · 3 linked inside Pith

  1. [1]

    Single-image HDR reconstruction by learning to reverse the camera pipeline,

    Y .-L. Liu, W.-S. Lai, Y .-S. Chen, Y .-L. Kao, M.-H. Yang, Y .-Y . Chuang, and J.-B. Huang, “Single-image HDR reconstruction by learning to reverse the camera pipeline,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1651–1660

    2020

  2. [2]

    Single-image hdr reconstruction by multi-exposure generation,

    P.-H. Le, Q. Le, R. Nguyen, and B.-S. Hua, “Single-image hdr reconstruction by multi-exposure generation,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2023, pp. 4063–4072

    2023

  3. [3]

    LiTMNet: A deep CNN for efficient HDR image reconstruction from a single LDR image,

    G. Wu, R. Song, M. Zhang, X. Li, and P. L. Rosin, “LiTMNet: A deep CNN for efficient HDR image reconstruction from a single LDR image,” Pattern Recognition, vol. 127, p. 108620, 2022

    2022

  4. [4]

    Lediff: Latent exposure diffusion for hdr generation,

    C. Wang, Z. Xia, T. Leimkuhler, K. Myszkowski, and X. Zhang, “Lediff: Latent exposure diffusion for hdr generation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 453– 464

    2025

  5. [5]

    Beyond feature mapping gap: Integrating real hdrtv priors for superior sdrtv-to-hdrtv conversion,

    G. He, K. Xu, L. Xu, S. Wang, W. Yu, and X. Wu, “Beyond feature mapping gap: Integrating real hdrtv priors for superior sdrtv-to-hdrtv conversion,”arXiv preprint arXiv:2411.10775, 2024

    arXiv 2024

  6. [6]

    Gain- MLP: Improving HDR Gain Map Encoding via a Lightweight MLP,

    T. D. Canham, S. Tedla, M. J. Murdoch, and M. S. Brown, “Gain- MLP: Improving HDR Gain Map Encoding via a Lightweight MLP,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 18619–18628

    2025

  7. [7]

    Learning gain map for inverse tone mapping,

    Y . Liao, Y . Guan, R. Xu, J. Li, S. Sun, and Z. Xiong, “Learning gain map for inverse tone mapping,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  8. [8]

    HDR Image Generation via Gain Map Decomposed Diffusion,

    Y . Guan, R. Xu, Y . Liao, M. Yao, L. Wang, and Z. Xiong, “HDR Image Generation via Gain Map Decomposed Diffusion,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 17536–17545

    2025

  9. [9]

    Retina inspired tone mapping method for high dynamic range images,

    X.-S. Zhang, K.-f. Yang, J. Zhou, and Y .-J. Li, “Retina inspired tone mapping method for high dynamic range images,”Optics express, vol. 28, no. 5, pp. 5953–5964, 2020

    2020

  10. [10]

    A biological vision inspired framework for image enhancement in poor visibility conditions,

    K.-F. Yang, X.-S. Zhang, and Y .-J. Li, “A biological vision inspired framework for image enhancement in poor visibility conditions,”IEEE Transactions on Image Processing, vol. 29, pp. 1493–1506, 2019

    2019

  11. [11]

    A biological retina inspired tone mapping processor for high-speed and energy-efficient image enhancement,

    X. Xiang, L. Liu, L. Que, C. Jia, B. Yan, Y . Li, J. Guo, and J. Zhou, “A biological retina inspired tone mapping processor for high-speed and energy-efficient image enhancement,”Sensors, vol. 20, no. 19, p. 5600, 2020

    2020

  12. [12]

    HDR image reconstruction from a single exposure using deep CNNs,

    G. Eilertsen, J. Kronander, G. Denes, R. K. Mantiuk, and J. Unger, “HDR image reconstruction from a single exposure using deep CNNs,” ACM transactions on graphics (TOG), vol. 36, no. 6, pp. 1–15, 2017

    2017

  13. [13]

    Single image HDR reconstruction using a CNN with masked features and perceptual loss,

    M. S. Santos, T. I. Ren, and N. K. Kalantari, “Single image HDR reconstruction using a CNN with masked features and perceptual loss,” arXiv preprint arXiv:2005.07335, 2020

    Pith/arXiv arXiv 2005

  14. [14]

    Deep reverse tone mapping,

    Y . Endo, Y . Kanamori, and J. Mitani, “Deep reverse tone mapping,”ACM Trans. Graph, vol. 36, no. 6, pp. 1–10, 2017

    2017

  15. [15]

    Hdrunet: Single image hdr reconstruction with denoising and dequantization,

    X. Chen, Y . Liu, Z. Zhang, Y . Qiao, and C. Dong, “Hdrunet: Single image hdr reconstruction with denoising and dequantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 354–363

    2021

  16. [16]

    A new journey from SDRTV to HDRTV ,

    X. Chen, Z. Zhang, J. S. Ren, L. Tian, Y . Qiao, and C. Dong, “A new journey from SDRTV to HDRTV ,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 4500–4509

    2021

  17. [17]

    Towards efficient sdrtv-to-hdrtv by learning from image formation,

    X. Chen, Z. Li, Z. Zhang, J. S. Ren, Y . Liu, J. He, Y . Qiao, J. Zhou, and C. Dong, “Towards efficient sdrtv-to-hdrtv by learning from image formation,”IEEE Transactions on Multimedia, vol. 27, pp. 8340 - 8354, 2025

    2025

  18. [18]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1833–1844

    2021

  19. [19]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17683–17693

    2022

  20. [20]

    Dual frequency transformer for efficient sdr-to-hdr translation,

    G. Xu, Q. Hou, and M.-M. Cheng, “Dual frequency transformer for efficient sdr-to-hdr translation,”Machine Intelligence Research, vol. 21, no. 3, pp. 538–548, 2024

    2024

  21. [21]

    Lightness and retinex theory,

    E. H. Land and J. J. McCann, “Lightness and retinex theory,”Journal of the Optical society of America, vol. 61, no. 1, pp. 1–11, 1971

    1971

  22. [22]

    A multiscale retinex for bridging the gap between color images and the human observation of scenes,

    D. J. Jobson, Z.-u. Rahman, and G. A. Woodell, “A multiscale retinex for bridging the gap between color images and the human observation of scenes,”IEEE Transactions on Image processing, vol. 6, no. 7, pp. 965–976, 1997

    1997

  23. [23]

    S-potentials from colour units in the retina of fish (Cyprinidae),

    K. I. Naka and W. A. Rushton, “S-potentials from colour units in the retina of fish (Cyprinidae),”The Journal of physiology, vol. 185, no. 3, pp. 536–555, 1966

    1966

  24. [24]

    A retinal adaptation model for HDR image compression,

    X. Pu, K. Yang, and Y . Li, “A retinal adaptation model for HDR image compression,” inCCF Chinese Conference on Computer Vision, 2017, pp. 37–47

    2017

  25. [25]

    A retina inspired model for high dynamic range image rendering,

    X.-S. Zhang and Y .-J. Li, “A retina inspired model for high dynamic range image rendering,” inInternational Conference on Brain Inspired Cognitive Systems, 2016, pp. 68–79

    2016

  26. [26]

    Retina-Inspired Models Enhance Visual Saliency Prediction,

    G. Shen, W. Ma, W. Zhai, X. Lv, G. Chen, and Y . Tian, “Retina-Inspired Models Enhance Visual Saliency Prediction,”Entropy, vol. 27, no. 4, p. 436, 2025. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, JUNE 2026 12

    2025

  27. [27]

    Color constancy using double-opponency,

    S.-B. Gao, K.-F. Yang, C.-Y . Li, and Y .-J. Li, “Color constancy using double-opponency,”IEEE transactions on pattern analysis and machine intelligence, vol. 37, no. 10, pp. 1973–1985, 2015

    1973

  28. [28]

    Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type,

    G.-q. Bi and M.-m. Poo, “Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type,”Journal of neuroscience, vol. 18, no. 24, pp. 10464–10472, 1998

    1998

  29. [29]

    Surrogate gradient learning in spiking neural networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 51–63, 2019

    2019

  30. [30]

    Spikformer: When spiking neural network meets transformer,

    Z. Zhou, Y . Zhu, C. He, Y . Wang, S. Yan, Y . Tian, and L. Yuan, “Spikformer: When spiking neural network meets transformer,”arXiv preprint arXiv:2209.15425, 2022

    Pith/arXiv arXiv 2022

  31. [31]

    Spike- driven transformer,

    M. Yao, J. Hu, Z. Zhou, L. Yuan, Y . Tian, B. Xu, and G. Li, “Spike- driven transformer,”Advances in neural information processing systems, vol. 36, pp. 64043–64058, 2023

    2023

  32. [32]

    Spiking neural networks for computer vision,

    M. Hopkins, G. Pineda-Garcia, P. A. Bogdan, and S. B. Furber, “Spiking neural networks for computer vision,”Interface focus, vol. 8, no. 4, 2018

    2018

  33. [33]

    High dynamic range imaging for dynamic scenes based on multi-level spike camera,

    Z. Zhu, R. Xiong, J. Zhao, R. Zhao, X. Fan, S. Zhu, and T. Huang, “High dynamic range imaging for dynamic scenes based on multi-level spike camera,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 6, pp. 5394–5406, 2025

    2025

  34. [34]

    Retina- inspired lightweight spiking convolutional neural network for single- image dehazing,

    Y . Zhang, X. Luo, Q. Sun, Y . Wang, H. Qu, and Z. Yi, “Retina- inspired lightweight spiking convolutional neural network for single- image dehazing,”IEEE Trans. Neural Netw. Learn. Syst., vol. 36, no. 7, pp. 12580–12592, 2025

    2025

  35. [35]

    PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR,

    M. Azimiet al., “PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR,” in2021 Picture Coding Symposium (PCS), 2021, pp. 1–5

    2021

  36. [36]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  37. [37]

    SR-SIM: A fast and high performance IQA index based on spectral residual,

    L. Zhang and H. Li, “SR-SIM: A fast and high performance IQA index based on spectral residual,” in2012 19th IEEE international conference on image processing, 2012, pp. 1473–1476

    2012

  38. [38]

    HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content,

    R. K. Mantiuk, D. Hammou, and P. Hanji, “HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and regular content,”arXiv preprint arXiv:2304.13625, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Physiological inverse tone mapping based on retina response,

    Y . Huo, F. Yang, L. Dong, and V . Brost, “Physiological inverse tone mapping based on retina response,”The Visual Computer, vol. 30, no. 5, pp. 507–517, 2014

    2014

  40. [40]

    High-quality reverse tone mapping for a wide range of exposures,

    R. P. Kovaleski and M. M. Oliveira, “High-quality reverse tone mapping for a wide range of exposures,” in2014 27th SIBGRAPI Conference on Graphics, Patterns and Images, 2014, pp. 49–56

    2014

  41. [41]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  42. [42]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1125– 1134

    2017

  43. [43]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 2223–2232

    2017

  44. [44]

    Deep bilateral learning for real-time image enhancement,

    M. Gharbi, J. Chen, J. T. Barron, S. W. Hasinoff, and F. Durand, “Deep bilateral learning for real-time image enhancement,”ACM Transactions on Graphics (TOG), vol. 36, no. 4, pp. 1–12, 2017

    2017

  45. [45]

    Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time,

    H. Zeng, J. Cai, L. Li, Z. Cao, and L. Zhang, “Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 4, pp. 2058–2073, 2020

    2058

  46. [46]

    Deep sr-itm: Joint learning of super- resolution and inverse tone-mapping for 4k uhd hdr applications,

    S. Y . Kim, J. Oh, and M. Kim, “Deep sr-itm: Joint learning of super- resolution and inverse tone-mapping for 4k uhd hdr applications,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 3116–3125

    2019

  47. [47]

    SI- HDR - dataset for comparison of single-image high dynamic range reconstruction methods,

    P. Hanji, R. Mantiuk, G. Eilertsen, S. Hajisharif, and J. Unger, “SI- HDR - dataset for comparison of single-image high dynamic range reconstruction methods,”Apollo - University of Cambridge Repository, 2022, doi: 10.17863/CAM.87333

    doi:10.17863/cam.87333 2022