pith. sign in

arxiv: 2606.05759 · v1 · pith:WM6H76CXnew · submitted 2026-06-04 · 💻 cs.CV

Physics-Guided Deep Unfolding for Blind Cross-Sensor Spectral Super-Resolution via Learning the Spectral Transformation Function

Zhaolin Li , Jinsong Chen , Shanxin Guo , Tuo Zhang , Xinglong Zhang , Pan Chen This is my paper

Pith reviewed 2026-06-28 02:53 UTC · model grok-4.3

classification 💻 cs.CV
keywords spectral super-resolutionblind estimationdeep unfoldinghyperspectral imagingspectral transformation functionphysics-guided networkcross-sensor
0
0 comments X

The pith

A deep unfolding network jointly recovers the hyperspectral image and the unknown spectral transformation function from multispectral inputs.

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

The paper targets blind cross-sensor spectral super-resolution, where the mapping from hyperspectral to multispectral images is unknown and changes with sensors and scenes. It introduces PGU-Net, which converts an alternating optimization loop into a fixed-stage trainable network. Each stage refines the hyperspectral estimate and the spectral transformation function by pairing learnable proximal networks with exact differentiable solvers. This structure is evaluated on CAVE and NTIRE 2022 benchmarks with varied spectral response functions plus a real UAV dataset pairing Headwall Nano and DJI P4 Multispectral sensors.

Core claim

PGU-Net unrolls an alternating optimization procedure into an end-to-end trainable architecture with stages, where each stage sequentially updates the HSI and the STF. Both modules combine learnable proximal networks with differentiable closed-form solvers, enabling physical interpretability while retaining strong representation capacity.

What carries the argument

PGU-Net, an unrolled alternating-optimization network that updates the hyperspectral image and the spectral transformation function in successive stages using proximal networks and closed-form solvers.

If this is right

  • The network produces both an improved HSI reconstruction and an explicit estimate of the unknown STF on datasets with multiple SRFs.
  • Performance holds under truly blind conditions on real cross-sensor UAV imagery without any ground-truth STF.
  • The estimated STF can capture land-cover-related spectral differences across scenes.
  • The method removes the need for sensor-specific calibration data that limits single-sensor SSR approaches.

Where Pith is reading between the lines

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

  • The same unrolling pattern could be applied to other blind imaging problems where the forward operator must be learned jointly with the signal.
  • Varying the number of stages at inference time might trade accuracy for speed without retraining.
  • The land-cover dependence noted in the STF estimate invites targeted experiments on homogeneous versus mixed scenes.

Load-bearing premise

The spectral degradation from HSI to MSI can be jointly estimated with the HSI itself through a fixed number of unrolled alternating optimization stages that combine learnable proximal networks with differentiable closed-form solvers.

What would settle it

On benchmark datasets supplied with known spectral response functions, the recovered STF deviates substantially from ground truth or the HSI reconstruction quality fails to exceed prior methods that assume a fixed known SRF.

Figures

Figures reproduced from arXiv: 2606.05759 by Jinsong Chen, Pan Chen, Shanxin Guo, Tuo Zhang, Xinglong Zhang, Zhaolin Li.

Figure 1
Figure 1. Figure 1: Spectral mixing process on (a) single-sensor scenario and (b) cross-sensors scenario [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the overall network. E. Overall PGU-Net Architectures The complete PGU-Net is formed by cascading 𝐾 stages, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Headwall–DJI cross-sensor UAV dataset. Left: false￾color composites (bands indicated by dashed lines) with a selected tree pixel (blue). Right: corresponding spectra from DJI MSI and Headwall HSI at the marked pixel. Compared Methods: In this study, we compare PGU-Net with representative SSR methods, including HSCNN+ [38], AWAN [39], MST++ [40] and SSR [26]. HSCNN+ and AWAN are CNN-based methods, MST++ is … view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of MSI reconstruction by STF predicted by PGU-Net on CAVE dataset. (1) MSI recovered by Nikon D700 SRF; (2) MSI recovered by STFs predicted by PGU￾Net; (3-5) RMSE error maps between (1) and (2) in red, green, and blue channels. (a-e) different samples named “Jellybeans”, “oil painting”, “paints”, “photo and face”, “pompoms” [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: A comparison of the SRFs recovered by the model (dashed line) with the ground truth (solid line); (a) CAVE dataset; (b) NTIRE 2022 dataset. To quantify the impact of these deviations, we use the estimated STF to degrade HSI and obtain a reconstructed MSI, which is then compared with the input MSI synthesized using the ground-truth SRF. Table I reports the quantitative results. The reconstructed MSIs are ne… view at source ↗
Figure 7
Figure 7. Figure 7: Recovered STFs (dashed lines) versus ground-truth SRFs (solid lines) for three given sensor types. (a) Landsat-7 ETM+. (b) Randomly generated Gaussian SRF. (c) Sentinel-2 MSI SRF. TABLE II HSI RECONSTRUCTION PERFORMANCE (MEAN ± STD. DEV.) ON CAVE DATASET WITH FOUR GIVEN SRFS. RMSE PSNR SSIM Nikon D700 0.0163± 0.0002 36.32 ± 0.11 0.9818 ± 0.0001 Landsat-7 0.0176 ± 0.0001 35.70 ± 0.04 0.9787 ± 0.0001 Gaussia… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of RGB reconstruction on NTIRE 2022: (1) MSI with ground-truth SRF; (2) MSI with SRFs recovered by model; (3-5) Error maps in red, green, and blue bands. (a-e) different samples named “ARAD_1K_0636”, “ARAD_1K_0671”, “ARAD_1K_0621”, “ARAD_1K_0945”, “ARAD_1K_0948”. We further evaluate STF estimation under three additional SRFs: Landsat-7 ETM+, randomly generated Gaussian SRFs, and Sentinel-2 MSI … view at source ↗
Figure 9
Figure 9. Figure 9: RMSE error map of different SSR methods comparison on (Up) the CAVE dataset “pompoms” ;(Down) NTIRE 2022 “ARAD_1K_0671”. TABLE III QUANTITATIVE COMPARISON OF DIFFERENT METHODS ON THE CAVE AND NTIRE 2022 DATASET CAVE NTIRE 2022 RMSE PSNR SSIM RMSE PSNR SSIM HSCNN+ 0.0494 26.18 0.9389 0.0588 26.36 0.9438 AWAN 0.0401 27.93 0.9550 0.0317 32.12 0.9646 MST++ 0.0281 31.02 0.9709 0.0247 34.32 0.9805 SSR 0.0194 34.… view at source ↗
Figure 11
Figure 11. Figure 11: presents the qualitative HSI reconstruction results. In [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: compares STFs estimated from (a) all pixels, (b) tree pixels, and (c) building pixels. The observed differences suggest that the effective STF may depend on scene content, which is consistent with the scene-dependent nature of spectral [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 12
Figure 12. Figure 12: RMSE maps of super-resolved HSI results from roof region (b3, b7) and road region (b6) in [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Consistency of the recovered STF across five random seeds. Gray curves: individual runs; shaded band: mean ±1 std; red curve: ground-truth SRF (Nikon D700). Table V FIVE INDEPENDENT RUNS WITH DIFFERENT RANDOM SEEDS (CAVE DATASET). Experiments RMSE PSNR SSIM Seed 42 0.0164 36.28 0.9816 Seed 123 0.0170 36.00 0.9797 Seed 1024 0.0167 36.18 0.9806 Seed 2025 0.0166 36.14 0.9809 Seed 5555 0.0170 35.99 0.9791 Mea… view at source ↗
read the original abstract

Hyperspectral imaging provides rich spectral information for quantitative remote sensing, yet hyperspectral sensors remain costly and thus unavailable in many UAV deployments. Spectral super-resolution (SSR) seeks to reconstruct hyperspectral images (HSIs) from multispectral images (MSIs). Most existing SSR methods assume a fixed and known spectral response function (SRF) and are therefore limited to single-sensor settings. In practical cross-sensor scenarios, the spectral degradation from HSI to MSI is unknown and varies with sensor characteristics and scene content, which renders HSI reconstruction ill-posed. This paper proposes a physics-guided deep unfolding network, termed PGU-Net, to address blind cross-sensor SSR by jointly estimating the HSI and a learnable spectral transformation function (STF). PGU-Net unrolls an alternating optimization procedure into an end-to-end trainable architecture with stages, where each stage sequentially updates the HSI and the STF. Both modules combine learnable proximal networks with differentiable closed-form solvers, enabling physical interpretability while retaining strong representation capacity. Experiments on benchmark datasets (CAVE and NTIRE 2022) with multiple SRFs demonstrate accurate recovery of the STF (degradation operator) and improved reconstruction performance over state-of-the-art SSR methods. Furthermore, evaluations on a real UAV cross-sensor dataset (Headwall Nano HSI and DJI P4 Multispectral MSI) verify the effectiveness and robustness of PGU-Net under truly blind conditions, and suggest that the estimated STF may exhibit land-cover-related differences.

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

2 major / 1 minor

Summary. The manuscript proposes PGU-Net, a physics-guided deep unfolding network for blind cross-sensor spectral super-resolution (SSR). It jointly estimates the hyperspectral image (HSI) and a learnable spectral transformation function (STF) by unrolling an alternating optimization procedure into an end-to-end trainable architecture. Each stage combines learnable proximal networks with differentiable closed-form solvers. Experiments on CAVE and NTIRE 2022 benchmarks with multiple SRFs, plus a real UAV cross-sensor dataset (Headwall Nano HSI and DJI P4 Multispectral MSI), are claimed to demonstrate accurate STF recovery and improved reconstruction over state-of-the-art SSR methods under blind conditions.

Significance. If the central claims hold, the work would address a practical gap in UAV-based hyperspectral imaging by enabling blind cross-sensor SSR without known SRFs. The physics-guided unfolding with closed-form solvers offers interpretability advantages over purely data-driven methods, and the joint STF/HSI estimation is a notable technical contribution for handling unknown degradation operators.

major comments (2)
  1. [Method description (alternating optimization unrolling)] The central assumption that fixed-stage unrolled alternating optimization (learnable proximal nets + differentiable closed-form solvers) can recover the true STF from MSI observations alone, without ground-truth STF or additional regularization on the STF, is load-bearing but insufficiently justified. Multiple (HSI, STF) pairs can produce identical MSI observations, raising identifiability risks that could lead the proximal networks to fit training-sensor artifacts rather than the underlying degradation; this directly impacts the claim of accurate STF recovery on the real UAV dataset where land-cover variation is noted as potentially affecting the estimated STF.
  2. [Experiments section] The experimental claims of 'accurate recovery of the STF' and 'improved reconstruction performance' on CAVE, NTIRE 2022, and the UAV dataset lack visible quantitative metrics, ablation studies, error bars, or full protocol details in the reported results, undermining verification of the central claims.
minor comments (1)
  1. Clarify the distinction (if any) between STF and SRF terminology throughout the manuscript.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and describe the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [Method description (alternating optimization unrolling)] The central assumption that fixed-stage unrolled alternating optimization (learnable proximal nets + differentiable closed-form solvers) can recover the true STF from MSI observations alone, without ground-truth STF or additional regularization on the STF, is load-bearing but insufficiently justified. Multiple (HSI, STF) pairs can produce identical MSI observations, raising identifiability risks that could lead the proximal networks to fit training-sensor artifacts rather than the underlying degradation; this directly impacts the claim of accurate STF recovery on the real UAV dataset where land-cover variation is noted as potentially affecting the estimated STF.

    Authors: We acknowledge the identifiability challenge inherent to blind STF estimation. The framework mitigates this through the physics-derived closed-form STF solver within each unfolding stage, which enforces consistency with the spectral degradation model, together with end-to-end training on diverse (HSI, MSI) pairs. On CAVE and NTIRE 2022 we validate recovered STFs against known ground-truth SRFs; on the UAV data the estimated STF yields measurable reconstruction gains even if land-cover effects are present. We will add an explicit discussion of identifiability, the role of the proximal networks as implicit regularizers, and the distinction between exact STF recovery and practical reconstruction utility. revision: partial

  2. Referee: [Experiments section] The experimental claims of 'accurate recovery of the STF' and 'improved reconstruction performance' on CAVE, NTIRE 2022, and the UAV dataset lack visible quantitative metrics, ablation studies, error bars, or full protocol details in the reported results, undermining verification of the central claims.

    Authors: The full manuscript contains quantitative tables (PSNR/SSIM/SAM for reconstruction and spectral error for STF) on CAVE and NTIRE 2022, plus qualitative UAV results. However, we agree that visibility, ablations, statistical reporting, and protocol details are insufficient. We will revise the experiments section to add: complete metric tables with all baselines, ablation studies on unfolding stages and STF module, error bars from repeated runs, and a detailed experimental protocol subsection covering hyperparameters, data splits, and training procedure. revision: yes

Circularity Check

0 steps flagged

No circularity; new trainable architecture with independent empirical claims

full rationale

The paper introduces PGU-Net as an end-to-end trainable deep unfolding network that jointly estimates HSI and STF via unrolled alternating optimization stages combining proximal networks and closed-form solvers. No derivation step reduces a claimed prediction or result to a fitted input by construction, nor does any load-bearing premise rely on self-citation chains or imported uniqueness theorems. The STF is explicitly an output of the learned model rather than presupposed, and performance claims rest on benchmark experiments and real UAV data rather than tautological reparameterization. This is a standard architecture proposal whose central content is independent of its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unrolling of alternating optimization being effective for joint estimation and on the differentiability of the closed-form solvers; these are domain assumptions rather than new physical entities. No free parameters beyond standard network weights are introduced in the abstract description.

free parameters (1)
  • Parameters of the learnable proximal networks
    Weights of the proximal networks are fitted during end-to-end training on the datasets.
axioms (1)
  • domain assumption The alternating optimization procedure for HSI and STF estimation can be unrolled into a fixed number of differentiable stages.
    The architecture is built by unrolling the alternating updates described in the abstract.

pith-pipeline@v0.9.1-grok · 5818 in / 1416 out tokens · 74363 ms · 2026-06-28T02:53:55.847359+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

26 extracted references · 22 canonical work pages

  1. [1]

    We formulate blind cross-sensor SSR as a joint estimation problem of the HSI X and an explicit spectral transformation matrix R (STF) that models the unknown cross-sensor spectral degradation

  2. [2]

    We develop PGU-Net, a physics-guided deep unfolding framework that unrolls an alternating optimization algorithm into a multi-stage network, integrating learnable proximal operators with differentiable closed-form solvers for both X and R

  3. [3]

    The remainder of this paper is organized as follows

    Extensive experiments on simulated and real cross-sensor data validate that PGU-Net improves reconstruction accuracy while yielding physically meaningful STF estimates; additional analyses indicate that the estimated STF exhibits consistent variations correlated with land-cover categories. The remainder of this paper is organized as follows. Section II pr...

  4. [4]

    Z-Net: The architecture employs a dual-branch design. The upper branch utilizes standard convolutions for global feature extraction, while the lower branch incorporates Channel Attention (CA) mechanisms to emphasize local spectral features. Features are fused via concatenation and refined through attention-enhanced convolutional layers to produce the regu...

  5. [5]

    Jellybeans

    P-Net: Taking the current estimate Rk as input, P-Net utilizes a multi-branch architecture with varying kernel sizes to extract multi-scale features. These are fused and processed by a transformer-based structure, which effectively models long-range dependencies and subtle spectral variations, yielding the regularized estimate Pk. Algorithm 1 The PGU-Net ...

    work page doi:10.3390/s19143071 2022

  6. [6]

    Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction,

    Y. Cai et al., “Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction,” Mar. 21, 2022, arXiv: arXiv:2111.07910. doi: 10.48550/arXiv.2111.07910

    work page doi:10.48550/arxiv.2111.07910 2022

  7. [7]

    ISPDiff: Interpretable Scale- Propelled Diffusion Model for Hyperspectral Image Super-Resolution,

    W. Dong, S. Liu, S. Xiao, J. Qu, and Y. Li, “ISPDiff: Interpretable Scale- Propelled Diffusion Model for Hyperspectral Image Super-Resolution,” Ieee T Geosci Remote, vol. 62, pp. 1–14, 2024, doi: 10.1109/TGRS.2024.3407967

    work page doi:10.1109/tgrs.2024.3407967 2024

  8. [8]

    EDiffSR: An Efficient Diffusion Probabilistic Model for Remote Sensing Image Super- Resolution,

    Y. Xiao, Q. Yuan, K. Jiang, J. He, X. Jin, and L. Zhang, “EDiffSR: An Efficient Diffusion Probabilistic Model for Remote Sensing Image Super- Resolution,” Ieee T Geosci Remote, vol. 62, pp. 1–14, 2024, doi: 10.1109/TGRS.2023.3341437

    work page doi:10.1109/tgrs.2023.3341437 2024

  9. [9]

    Coupling model- and data-driven methods for remote sensing image restoration and fusion: Improving physical interpretability,

    H. Shen, M. Jiang, J. Li, C. Zhou, Q. Yuan, and L. Zhang, “Coupling model- and data-driven methods for remote sensing image restoration and fusion: Improving physical interpretability,” IEEE Geoscience and Remote Sensing Magazine, vol. 10, no. 2, pp. 231–249, June 2022, doi: 10.1109/MGRS.2021.3135954

    work page doi:10.1109/mgrs.2021.3135954 2022

  10. [10]

    Spectral Response Function-Guided Deep Optimization-Driven Network for Spectral Super- Resolution,

    J. He, J. Li, Q. Yuan, H. Shen, and L. Zhang, “Spectral Response Function-Guided Deep Optimization-Driven Network for Spectral Super- Resolution,” Ieee T Neur Net Lear, vol. 33, no. 9, pp. 4213–4227, Sept. 2022, doi: 10.1109/TNNLS.2021.3056181

    work page doi:10.1109/tnnls.2021.3056181 2022

  11. [11]

    Spectral Super-Resolution via Model-Guided Cross-Fusion Network,

    R. Dian, T. Shan, W. He, and H. Liu, “Spectral Super-Resolution via Model-Guided Cross-Fusion Network,” Ieee T Neur Net Lear, pp. 1–12, 2023, doi: 10.1109/TNNLS.2023.3238506

    work page doi:10.1109/tnnls.2023.3238506 2023

  12. [12]

    Filter Selection for Hyperspectral Estimation,

    B. Arad and O. Ben-Shahar, “Filter Selection for Hyperspectral Estimation,” in 2017 IEEE International Conference on Computer Vision (ICCV), Oct. 2017, pp. 3172–3180. doi: 10.1109/ICCV.2017.342

    work page doi:10.1109/iccv.2017.342 2017

  13. [13]

    Joint Camera Spectral Response Selection and Hyperspectral Image Recovery,

    Y. Fu, T. Zhang, Y. Zheng, D. Zhang, and H. Huang, “Joint Camera Spectral Response Selection and Hyperspectral Image Recovery,” Ieee T Pattern Anal, vol. 44, no. 1, pp. 256–272, Jan. 2022, doi: 10.1109/TPAMI.2020.3009999

    work page doi:10.1109/tpami.2020.3009999 2022

  14. [14]

    Efficient transfer learning for spectral image reconstruction from RGB images,

    E. Martínez, S. Castro, J. Bacca, and H. Arguello, “Efficient transfer learning for spectral image reconstruction from RGB images,” in 2020 IEEE Colombian Conference on Applications of Computational Intelligence (IEEE ColCACI 2020), Aug. 2020, pp. 1–6. doi: 10.1109/ColCACI50549.2020.9247895

    work page doi:10.1109/colcaci50549.2020.9247895 2020

  15. [15]

    Progressive Spatial–Spectral Joint Network for Hyperspectral Image Reconstruction,

    T. Li and Y. Gu, “Progressive Spatial–Spectral Joint Network for Hyperspectral Image Reconstruction,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2022, doi: 10.1109/TGRS.2021.3079969

    work page doi:10.1109/tgrs.2021.3079969 2022

  16. [16]

    Globally quantitative analysis of the impact of atmosphere and spectral response function on 2-band enhanced vegetation index (EVI2) over sentinel-2 and landsat-8,

    Z. Zhen, S. Chen, T. Yin, and J.-P. Gastellu-Etchegorry, “Globally quantitative analysis of the impact of atmosphere and spectral response function on 2-band enhanced vegetation index (EVI2) over sentinel-2 and landsat-8,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 205, pp. 206–226, Nov. 2023, doi: 10.1016/j.isprsjprs.2023.09.024

    work page doi:10.1016/j.isprsjprs.2023.09.024 2023

  17. [17]

    Blind Spectral Super- Resolution by Estimating Spectral Degradation Between Unpaired Images,

    J. Xie, L. Fang, C. Wu, F. Xie, and J. Chanussot, “Blind Spectral Super- Resolution by Estimating Spectral Degradation Between Unpaired Images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–14, 2024, doi: 10.1109/TGRS.2024.3387857

    work page doi:10.1109/tgrs.2024.3387857 2024

  18. [18]

    Advancing image super-resolution techniques in remote sensing: A comprehensive survey,

    Y. Qi, M. Lou, Y. Liu, L. Li, Z. Yang, and W. Nie, “Advancing image super-resolution techniques in remote sensing: A comprehensive survey,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 231, pp. 68– 100, Jan. 2026, doi: 10.1016/j.isprsjprs.2025.10.024

    work page doi:10.1016/j.isprsjprs.2025.10.024 2026

  19. [19]

    Physics-informed hyperspectral remote sensing image synthesis with deep conditional generative adversarial networks,

    L. Liu, W. Li, Z. Shi, and Z. Zou, “Physics-informed hyperspectral remote sensing image synthesis with deep conditional generative adversarial networks,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–15, 2022, doi: 10.1109/TGRS.2022.3173532

    work page doi:10.1109/tgrs.2022.3173532 2022

  20. [20]

    Single Hyperspectral Image Super-Resolution with Grouped Deep Recursive Residual Network,

    Y. Li, L. Zhang, C. Dingl, W. Wei, and Y. Zhang, “Single Hyperspectral Image Super-Resolution with Grouped Deep Recursive Residual Network,” in 2018 IEEE Fourth International Conference on Multimedia Big Data (BigMM) , Sept. 2018, pp. 1–4. doi: 10.1109/BigMM.2018.8499097

    work page doi:10.1109/bigmm.2018.8499097 2018

  21. [21]

    Generalized Assorted Pixel Camera: Postcapture Control of Resolution, Dynamic Range, and Spectrum,

    F. Yasuma, T. Mitsunaga, D. Iso, and S. K. Nayar, “Generalized Assorted Pixel Camera: Postcapture Control of Resolution, Dynamic Range, and Spectrum,” IEEE Trans. on Image Process., vol. 19, no. 9, pp. 2241–2253, Sept. 2010, doi: 10.1109/TIP.2010.2046811

    work page doi:10.1109/tip.2010.2046811 2010

  22. [22]

    Roadsaw: A large-scale dataset for camera- based road surface and wetness estimation,

    B. Arad et al., “NTIRE 2022 spectral recovery challenge and data set,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June 2022, pp. 862–880. doi: 10.1109/CVPRW56347.2022.00102

    work page doi:10.1109/cvprw56347.2022.00102 2022

  23. [23]

    HSCNN+: Advanced CNN-Based Hyperspectral Recovery from RGB Images,

    Z. Shi, C. Chen, Z. Xiong, D. Liu, and F. Wu, “HSCNN+: Advanced CNN-Based Hyperspectral Recovery from RGB Images,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June 2018, pp. 1052–10528. doi: 10.1109/CVPRW.2018.00139

    work page doi:10.1109/cvprw.2018.00139 2018

  24. [24]

    Frerix, T., Niesner, M., and Cremers, D

    J. Li, C. Wu, R. Song, Y. Li, and F. Liu, “Adaptive Weighted Attention Network with Camera Spectral Sensitivity Prior for Spectral Reconstruction from RGB Images,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June 2020, pp. 1894–1903. doi: 10.1109/CVPRW50498.2020.00239

    work page doi:10.1109/cvprw50498.2020.00239 2020

  25. [25]

    Roadsaw: A large-scale dataset for camera- based road surface and wetness estimation,

    Y. Cai et al., “MST++: Multi-stage Spectral-wise Transformer for Efficient Spectral Reconstruction,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June 2022, pp. 744–754. doi: 10.1109/CVPRW56347.2022.00090

    work page doi:10.1109/cvprw56347.2022.00090 2022

  26. [26]

    Introduction to PyTorch,

    N. Ketkar, “Introduction to PyTorch,” in Deep Learning with Python: A Hands-on Introduction, N. Ketkar, Ed., Berkeley, CA: Apress, 2017, pp. 195–208. doi: 10.1007/978-1-4842-2766-4_12

    work page doi:10.1007/978-1-4842-2766-4_12 2017