Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Unfolding Framework with Complex-Valued Deformable Attention for High-Quality Computer-Generated Hologram Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a deep unfolding network, which alternates a learned adaptive bandwidth-preserving diffraction model with a complex-valued denoiser using deformable self-attention, generates computer-generated holograms at 1920×1080

desk verdict The paper's core gradient update has a sign error that breaks the deep-unfolding claim, but the architecture and results are worth a hard look if the code confirms the intended sign. read the letter →

arxiv 2508.21657 v1 pith:HZJ75TFX submitted 2025-08-29 cs.CV

Haomiao Zhang , Zhangyuan Li , Yanling Piao , Zhi Li , Xiaodong Wang , Miao Cao , Xiongfei Su , Qiang Song
show 1 more author
Xin Yuan
This is my paper · ORCID
classification cs.CV PACS 42.40.Jf
keywords computer-generatedholographydeepunfoldingnetworkcomplex-valueddeformableattentionadaptivebandwidth-preservingmethodphaseretrievalholographicdisplay
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes replacing the black-box end-to-end paradigm for computer-generated holography with a deep unfolding network that explicitly alternates between a physics-based propagation step and a learned denoising step. The propagation step uses a new adaptive bandwidth-preserving method (ABPM) that extends accurate diffraction calculations beyond the 23.62 cm limit of the angular spectrum method. The denoising step uses a phase-domain complex-valued denoiser with deformable self-attention, which captures global phase relationships cheaply. The authors report state-of-the-art reconstruction quality on simulated and real optical setups, with a PSNR improvement of 4.44 to 15.19 dB over CNN-based baselines. The approach matters because it offers a flexible, interpretable alternative that works across near and far fields without retraining when optical parameters change.

What carries the argument

The key machinery is the ABPM propagation operator, a three-regime diffraction model that switches between direct angular-spectrum, impulse-response, and far-field impulse-response forms based on sampling thresholds z1 and z2, plus the complex-valued deformable self-attention (CDSA) in the denoiser. ABPM is what extends the working distance beyond the ASM threshold while keeping spatial sampling fixed. CDSA reduces the attention matrix size by 64× through deformable down-sampling and uses the Hermitian inner product ⟨Q,K⟩ so that phase differences, not just amplitudes, drive the attention weights; this preserves global context at a practical computational cost.

What would settle it

Train the model on DIV2K and evaluate on a rigorously disjoint test set; if the average PSNR falls below 35 dB or the ABPM reconstruction degrades beyond 23.62 cm once test/train overlap is removed, the state-of-the-art and wide-distance claims would be unsupported. Alternatively, re-derive the sampling thresholds in Eq. (6) and verify whether the impulse-response sampling satisfies the Nyquist criterion at 60 cm; if it does not, the far-field accuracy claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that the ill-posed phase-retrieval problem of computer-generated holography, formulated as y = |Φx|, can be solved by unrolling gradient descent into two alternating modules: an adaptive bandwidth-preserving propagation model (ABPM) that keeps sampling within Nyquist bounds over a wide range of distances, and a Phase-domain Complex-valued Denoiser (PCD) whose attention mechanism respects complex-valued phase structure. The paper shows that ABPM, with distance-dependent sampling thresholds from Eq. (6), preserves image-plane pixel size and avoids spectral aliasing far beyond the standard ASM threshold. The PCD's complex-valued deformable self-attention down-samples keys a

Load-bearing premise

The reported quality numbers assume the 100 validation images were not also used in training, but the paper never states that the test set is disjoint from the 800 training images.

Editorial extensions

If this is right

  • If the paper is right, holographic display systems can use a single trained network across propagation distances from 8 to 60 cm, avoiding the need to retrain for each working distance.
  • The deep unfolding structure allows optical parameters such as wavelength or distance to be changed at inference by swapping the ABPM forward model, giving a flexibility that end-to-end black-box networks lack.
  • The reported 36.45 dB PSNR at 1920×1080 indicates that physics-informed unfolding outperforms pure CNN and traditional iterative methods, suggesting that learned priors combined with explicit forward models are a practical path to high-quality CGH.
  • The complex-valued deformable self-attention mechanism with Hermitian inner product provides a scalable way to capture global phase relationships, enabling attention-based CGH at high resolution with low memory and FLOPs.
  • The method's success in real optical experiments implies that ABPM is not merely a simulation trick, but works with an actual SLM and camera setup.

Reading between the lines

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

  • The same unfolding architecture could be transferred to other coherent imaging problems that rely on angular-spectrum-like propagation, such as in-line holography, Fourier ptychography, or diffractive tomography, where distance-dependent bandwidth preservation is equally critical.
  • The paper's ablation shows that increasing embedding channels from 32 to 96 raises PSNR from 36.45 to 36.87 dB, hinting that larger models offer headroom for even higher quality if compute allows, a direction the paper does not pursue.
  • Since training is done on natural images from DIV2K, the model may be biased toward photographic content; testing on text, synthetic scenes, or dense speckle patterns would reveal whether the learned prior generalizes beyond natural-image statistics.
  • The Hermitian-inner-product attention is a generic complex-valued building block that could benefit other complex-domain tasks, including MRI phase reconstruction and complex-valued communication signal processing, outside of holography.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a deep-unfolding network (DUN) for computer-generated holography. The forward model is y=|Φx|. The DUN alternates a gradient-descent update (Eq. 4) using an adaptive bandwidth-preserving propagation model (ABPM) and a learned phase-domain complex-valued denoiser (PCD) with complex-valued deformable self-attention (CDSA). Experiments on simulated 1920×1080 images report state-of-the-art PSNR 36.45 dB / SSIM 0.95, and optical experiments are shown at distances 10–60 cm. The central claim is that decomposing gradient descent into ABPM+PCD gives interpretable, flexible, high-quality CGH, with ABPM extending working distance beyond the ASM threshold.

Significance. If the claims hold, the work is a meaningful step: it provides a physics-grounded deep-unfolding architecture for CGH, a bandwidth-preserving propagation formulation that extends the working distance, and a memory-efficient complex-valued deformable attention mechanism. The paper includes a code link, an ablation of the denoiser, and a large reported PSNR gain over its three baselines. However, the significance is contingent on correcting a central gradient-sign issue and on a clean evaluation protocol; without those, the method is not demonstrably a deep unfolding of the stated objective, and the state-of-the-art claim is not yet established.

major comments (4)
  1. [Sec. IV-A, Eq. (4)] The sign in the gradient update is wrong. For F(x)=1/2||y−|Φx|||², the Wirtinger derivative is ∂F/∂x* = (1/2)Φ^H diag(Φx/|Φx|)(|Φx|−y), so a gradient-descent step is x + (ρ/2)Φ^H diag(Φx/|Φx|)(y−|Φx|). Eq. (4) as written, x − Φ^H diag(...)(y−|Φx|), has the opposite sign and implements gradient ascent, not descent. This invalidates the claim that Eq. (4) is the closed-form solution of Eq. (3a) and the interpretation of the module as unfolding gradient descent. If the released code uses the plus sign, Eq. (4) is a serious typo that must be fixed; if the code uses the minus sign, the physics-driven unfolding claim collapses. Please correct the equation and, ideally, verify the implemented update.
  2. [Sec. V-A] The training/evaluation split is not stated clearly. The text says the model is trained on DIV2K (800 images) and evaluated on 100 randomly selected images from DIV2K and Flickr2K. DIV2K consists of 800 training images plus a separate validation set. If any of the 100 test images are drawn from the 800 training images, the reported PSNR/SSIM values are inflated by train/test leakage and the comparisons in Table I are invalid. The paper must state explicitly whether the test images are disjoint from the training set, or re-run the evaluation on a held-out set.
  3. [Sec. V-C] The real-data results are presented only qualitatively in Fig. 5. The text claims 'superior reconstruction performance' and validates wide working distance, but no PSNR, SSIM, or user study is reported on the optical data. Because the real-data demonstration is a stated contribution, quantitative metrics (or an explicit statement that no quantitative optical evaluation was performed) are needed to support the claim.
  4. [Sec. V-B and Table I] The 'state-of-the-art' claim is not supported by the chosen baselines. Table I compares only GS, HoloNet (2020), and CCNN-CGH (2023). Several more recent CGH methods mentioned in the references (e.g., 4K-DMDNet, camera-in-the-loop, Tensor Holography-style approaches) are not compared. Either add recent baselines under the same experimental conditions or temper the state-of-the-art claim to 'outperforms the tested baselines'.
minor comments (5)
  1. [Sec. IV-A, Eq. (4)] Even setting aside the sign, the coefficient is inconsistent: the text says the 1/2 is absorbed into ρ and ρ=1, but the second equality drops ρ entirely. Please clarify the exact relation between ∇F and the displayed update.
  2. [Sec. IV-A] The phrase 'where ∇xF(x(k)) denotes the gradient of F(x) = |Φx|' is mathematically imprecise: F(x) is a scalar objective, not the vector |Φx|. Please fix the notation.
  3. [Sec. IV-D] The sentence 'Since the similarity dominates the attention matrix (due to initially small values in T)' is unclear. What is T? Please rephrase.
  4. [Table II] The CVTF baseline is evaluated at 256×256 resolution while other rows are at 1920×1080. This should be stated in the table caption or text so the comparison is not misleading.
  5. [Table III] The embedding-channel ablation shows C=96 gives 36.87 dB versus 36.45 dB for C=32, at roughly 4.5× the parameters and 10× the FLOPs. The text says C=32 'achieves excellent results,' but the small accuracy gain from C=96 should be discussed to justify the lightweight choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the unfolding derivation, ABPM sampling thresholds, and CDSA attention are derived from external optimization and sampling theory, and the self-citations are not load-bearing.

full rationale

The paper's central derivation chain is self-contained. The deep-unfolding update in Eq. (4) is presented as a gradient step for the phase-retrieval fidelity term; whether the sign is correct is a correctness concern, not a circularity, because the update is not defined in terms of the quantity it later predicts. The ABPM model in Eq. (6) is justified by sampling-theoretic thresholds (z1, z2) whose derivation is deferred to the supplementary material, but nothing in the main text indicates these were fitted to the reported PSNR values. The PCD denoiser is learned on DIV2K and evaluated on '100 randomly selected images from the DIV2K and Flickr2K datasets'; if the test set overlapped the training set, the reported PSNR would be a fit-quality measure rather than a free prediction, but the paper does not state such overlap and common dataset conventions imply disjoint splits, so no circular reduction can be exhibited. The self-citations [27], [28] support the generic statement that physics-driven DUNs can outperform E2E networks in some tasks; this claim is not load-bearing because the paper's own comparisons in Table I provide the primary evidence for the method's performance. No uniqueness theorem is imported from the authors' prior work, no ansatz is smuggled in via citation, and ABPM is a renamed combination of existing angular-spectrum sampling ideas but is not presented as a derivation of a new physical law. Therefore no step reduces by construction to its own inputs.

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

The central claim rests on standard phase retrieval assumptions and a learned image prior. No ad hoc physical constants are introduced for the forward model; the network weights are trained, and architecture choices are listed as free parameters for transparency.

free parameters (4)
  • Gradient step size rho = 1
    Set in Eq. (4); step size of the unrolled gradient descent.
  • Number of unfolding stages N
    Not reported in the main text; only diagrammed as N stages in Fig. 2.
  • CDSA down-sample rate and kernel = 64, 9x9
    Chosen for efficiency; controls attention region density.
  • Embedding channels C = 32
    Selected in ablation to balance params and PSNR.
assumptions (4)
  • domain assumption Measurement is closer to amplitude |Uz| than intensity |Uz|^2
    Used to define the forward model y = |Phi x| in Sec. III, following [23, 29].
  • standard math Half quadratic splitting decouples fidelity and prior, and a denoiser acts as the proximal operator
    Used in Sec. IV-A to justify Eqs. (3)-(5), following [30].
  • standard math The gradient of the fidelity term has the closed form given in Eq. (4)
    Assumed from [29]; sign conventions are absorbed by the learned step.
  • domain assumption ASM sampling thresholds z1 and z2 define valid propagation regimes
    Eq. (6) relies on sampling theory from [9-12]; derivation is deferred to the SM.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unfolding Framework with Complex-Valued Deformable Attention for High-Quality Computer-Generated Hologram Generation." pith.science (2026). https://pith.science/paper/HZJ75TFX

@misc{pith2026250821657,
  author       = {Pith},
  title        = {Pith review of: Unfolding Framework with Complex-Valued Deformable Attention for High-Quality Computer-Generated Hologram Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZJ75TFX}},
  note         = {Machine review of arXiv:2508.21657}
}
abstract

Computer-generated holography (CGH) has gained wide attention with deep learning-based algorithms. However, due to its nonlinear and ill-posed nature, challenges remain in achieving accurate and stable reconstruction. Specifically, ($i$) the widely used end-to-end networks treat the reconstruction model as a black box, ignoring underlying physical relationships, which reduces interpretability and flexibility. ($ii$) CNN-based CGH algorithms have limited receptive fields, hindering their ability to capture long-range dependencies and global context. ($iii$) Angular spectrum method (ASM)-based models are constrained to finite near-fields.In this paper, we propose a Deep Unfolding Network (DUN) that decomposes gradient descent into two modules: an adaptive bandwidth-preserving model (ABPM) and a phase-domain complex-valued denoiser (PCD), providing more flexibility. ABPM allows for wider working distances compared to ASM-based methods. At the same time, PCD leverages its complex-valued deformable self-attention module to capture global features and enhance performance, achieving a PSNR over 35 dB. Experiments on simulated and real data show state-of-the-art results.

Figures

Figures reproduced from arXiv: 2508.21657 by the authors.

Figure 1
Figure 1. (a) Schematic of CGH model. (b) Experiment setup of our CGH system (BS: beam splitter, CCD: charge-coupled device, P: polarizer). (c) Comparison between the reconstruction results of several representative CGH algorithms in 256×256 resolution. challenges in capturing long-range relationships. In contrast, Transformer models excel at capturing long-range, non-local relationships, improving reconstruction quality by p… view at source ↗
Figure 2
Figure 2. The overall framework of our method, containing a deep-unfolding structure with N stages; GD and PCD blocks represent the operation in Eq.(4) and Eq.(5) respectively. age features and physical constraints in CGH reconstruc￾tion, which consists of a gradient descent (GD) module and a deep denoising module to enhance both flexibility and interpretability. • In the GD module, we replace the widely-used ASM model with a… view at source ↗
Figure 3
Figure 3. (a) The diagram of CDAT block. (b) Complex-valued Deformable Self-Attention (CDSA) module. (c) An illustration of the difference between Hermitian inner production and dot production. where ∇xF(x (k) ) denotes the gradient of F(x) = |Φx| at the point x = x (k) . ∇xF(x (k) ) is derived as ∇xF(x) = 1 2Φ Hdiag  Φx |Φx|  (y − |Φx|), where (·) H denotes the con￾jugate transpose operator. Notice that 1 2 is absorbed int… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Simulation:(a) Numerical reconstructions of CGH generated by different algorithms on four benchmark datasets at 1920×1080 resolution. (b) CGH reconstruction results of our algorithm at 8 cm, 20 cm, 40 cm, and 60 cm. Zoom in for a better view. of 64. The offset is gener…
Figure 5
Figure 5. Figure 5: Real data results: Reconstructed results of various algorithms at different propagation distances with ASM threshold. (b) Reconstructed results of our algorithm beyond the ASM threshold. Zoom in for a better view. We compare the performance of various algorithms in the…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 31 canonical work pages

  1. [1]

    A practical algorithm for the determination of plane from image and diffraction pictures,

    R. W. Gerchberg, “A practical algorithm for the determination of plane from image and diffraction pictures,” Optik, vol. 35, no. 2, pp. 237–246, 1972

    1972

  2. [2]

    Gradient descent with random initialization: Fast global convergence for nonconvex phase retrieval,

    Y . Chen, Y . Chi, J. Fan, and C. Ma, “Gradient descent with random initialization: Fast global convergence for nonconvex phase retrieval,” Mathematical Programming, vol. 176, pp. 5–37, 2019

    work page 2019

  3. [3]

    Phase retrieval via wirtinger flow: Theory and algorithms,

    E. J. Cand `es, X. Li, and M. Soltanolkotabi, “Phase retrieval via wirtinger flow: Theory and algorithms,” IEEE Transactions on Information The- ory, vol. 61, no. 4, pp. 1985–2007, 2015

    work page 1985

  4. [4]

    Diffraction model-informed neural network for unsupervised layer-based computer- generated holography,

    X. Shui, H. Zheng, X. Xia, F. Yang, W. Wang, and Y . Yu, “Diffraction model-informed neural network for unsupervised layer-based computer- generated holography,” Optics Express , vol. 30, no. 25, pp. 44 814– 44 826, 2022

    work page 2022

  5. [5]

    4k-dmdnet: diffraction model- driven network for 4k computer-generated holography,

    K. Liu, J. Wu, Z. He, and L. Cao, “4k-dmdnet: diffraction model- driven network for 4k computer-generated holography,” Opto-Electronic Advances, pp. 220 135–1, 2023

    work page 2023

  6. [6]

    Learning fast approximations of sparse coding,

    K. Gregor and Y . LeCun, “Learning fast approximations of sparse coding,” in Proceedings of the 27th international conference on inter- national conference on machine learning , 2010, pp. 399–406

    work page 2010

  7. [7]

    Embedding fourier for ultra-high-definition low-light image enhance- ment,

    C. Li, C.-L. Guo, M. Zhou, Z. Liang, S. Zhou, R. Feng, and C. C. Loy, “Embedding fourier for ultra-high-definition low-light image enhance- ment,” arXiv preprint arXiv:2302.11831 , 2023

    arXiv 2023

  8. [8]

    Fourmer: An efficient global modeling paradigm for image restoration,

    M. Zhou, J. Huang, C.-L. Guo, and C. Li, “Fourmer: An efficient global modeling paradigm for image restoration,” International conference on machine learning , pp. 42 589–42 601, 2023

    work page 2023

Show all 34 references
  1. [9]

    J. W. Goodman, Introduction to F ourier optics. Roberts and Company publishers, 2005

    2005

  2. [10]

    Band-limited angular spectrum method for numerical simulation of free-space propagation in far and near fields,

    K. Matsushima and T. Shimobaba, “Band-limited angular spectrum method for numerical simulation of free-space propagation in far and near fields,” Optics express, vol. 17, no. 22, pp. 19 662–19 673, 2009

    2009

  3. [11]

    Non-uniform sampling and wide range angular spectrum method,

    Y .-H. Kim, C.-W. Byun, H. Oh, J. Lee, J.-E. Pi, G. H. Kim, M.-L. Lee, H. Ryu, H.-Y . Chu, and C.-S. Hwang, “Non-uniform sampling and wide range angular spectrum method,” Journal of Optics , vol. 16, no. 12, p. 125710, 2014

    2014

  4. [12]

    Adaptive-sampling angular spectrum method with full utilization of space-bandwidth product,

    W. Zhang, H. Zhang, and G. Jin, “Adaptive-sampling angular spectrum method with full utilization of space-bandwidth product,” Optics Letters, vol. 45, no. 16, pp. 4416–4419, 2020

    2020

  5. [13]

    Signal transformer: Complex-valued attention and meta-learning for signal recognition,

    Y . Dong, Y . Peng, M. Yang, S. Lu, and Q. Shi, “Signal transformer: Complex-valued attention and meta-learning for signal recognition,” arXiv preprint arXiv:2106.04392 , 2021

    2021 arXiv

  6. [14]

    Building blocks for a complex-valued trans- former architecture,

    F. Eilers and X. Jiang, “Building blocks for a complex-valued trans- former architecture,” in ICASSP 2023-2023 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

    2023

  7. [15]

    Complex transformer: A framework for modeling complex-valued sequence,

    M. Yang, M. Q. Ma, D. Li, Y .-H. H. Tsai, and R. Salakhutdinov, “Complex transformer: A framework for modeling complex-valued sequence,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2020, pp. 4232–4236

    2020

  8. [16]

    Review of fast methods for point- based computer-generated holography,

    P. Tsang, T.-C. Poon, and Y . Wu, “Review of fast methods for point- based computer-generated holography,” Photonics Research , vol. 6, no. 9, pp. 837–846, 2018

    2018

  9. [17]

    Holographic near-eye displays for virtual and augmented reality,

    A. Maimone, A. Georgiou, and J. S. Kollin, “Holographic near-eye displays for virtual and augmented reality,” ACM Transactions on Graphics (Tog), vol. 36, no. 4, pp. 1–16, 2017

    2017

  10. [18]

    Deep learning for hologram generation,

    S.-C. Liu and D. Chu, “Deep learning for hologram generation,” Optics Express, vol. 29, no. 17, pp. 27 373–27 395, 2021

    2021

  11. [19]

    Towards real- time photorealistic 3d holography with deep neural networks,

    L. Shi, B. Li, C. Kim, P. Kellnhofer, and W. Matusik, “Towards real- time photorealistic 3d holography with deep neural networks,” Nature, vol. 591, no. 7849, pp. 234–239, 2021

    2021

  12. [20]

    End-to-end learning of 3d phase- only holograms for holographic display,

    L. Shi, B. Li, and W. Matusik, “End-to-end learning of 3d phase- only holograms for holographic display,” Light: Science & Applications , vol. 11, no. 1, p. 247, 2022

    2022

  13. [21]

    From picture to 3d hologram: end-to-end learning of real-time 3d photorealistic hologram generation from 2d image input,

    C. Chang, B. Dai, D. Zhu, J. Li, J. Xia, D. Zhang, L. Hou, and S. Zhuang, “From picture to 3d hologram: end-to-end learning of real-time 3d photorealistic hologram generation from 2d image input,” Optics Letters, vol. 48, no. 4, pp. 851–854, 2023

    2023

  14. [22]

    High-speed computer-generated holography using an autoencoder-based deep neural network,

    J. Wu, K. Liu, X. Sui, and L. Cao, “High-speed computer-generated holography using an autoencoder-based deep neural network,” Optics Letters, vol. 46, no. 12, pp. 2908–2911, 2021

    2021

  15. [23]

    Neural hologra- phy with camera-in-the-loop training,

    Y . Peng, S. Choi, N. Padmanaban, and G. Wetzstein, “Neural hologra- phy with camera-in-the-loop training,” ACM Transactions on Graphics (TOG), vol. 39, no. 6, pp. 1–14, 2020

    2020

  16. [24]

    Complex-domain-enhancing neural network for large-scale coherent imaging,

    X. Chang, R. Zhao, S. Jiang, C. Shen, G. Zheng, C. Yang, and L. Bian, “Complex-domain-enhancing neural network for large-scale coherent imaging,” Advanced Photonics Nexus, vol. 2, no. 4, pp. 046 006–046 006, 2023

    2023

  17. [25]

    Real-time high-quality computer-generated hologram using complex-valued convolutional neural network,

    C. Zhong, X. Sang, B. Yan, H. Li, D. Chen, X. Qin, S. Chen, and X. Ye, “Real-time high-quality computer-generated hologram using complex-valued convolutional neural network,” IEEE Transactions on Visualization and Computer Graphics , 2023

    2023

  18. [26]

    Vision transformer- based, high-fidelity, computer-generated holography,

    Z. Dong, C. Xu, Y . Tang, Y . Ling, Y . Li, and Y . Su, “Vision transformer- based, high-fidelity, computer-generated holography,” in Advances in Display Technologies XIII , vol. 12443. SPIE, 2023, pp. 47–53

    2023

  19. [27]

    Physics-driven deep learning enables temporal compressive coherent diffraction imaging,

    Z. Chen, S. Zheng, Z. Tong, and X. Yuan, “Physics-driven deep learning enables temporal compressive coherent diffraction imaging,” Optica, vol. 9, no. 6, pp. 677–680, 2022

    2022

  20. [28]

    Dual-scale transformer for large-scale single-pixel imaging,

    G. Qu, P. Wang, and X. Yuan, “Dual-scale transformer for large-scale single-pixel imaging,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 25 327–25 337

    2024

  21. [29]

    Iterative projection meets sparsity regularization: towards practical single-shot quantitative phase imaging with in-line holography,

    Y . Gao and L. Cao, “Iterative projection meets sparsity regularization: towards practical single-shot quantitative phase imaging with in-line holography,” Light: Advanced Manufacturing , vol. 4, no. 1, pp. 1–17, 2023

    2023

  22. [30]

    From denoising to compressed sensing,

    C. A. Metzler, A. Maleki, and R. G. Baraniuk, “From denoising to compressed sensing,” IEEE Transactions on Information Theory, vol. 62, no. 9, pp. 5117–5144, 2016

    2016

  23. [31]

    Vision transformer with deformable attention,

    Z. Xia, X. Pan, S. Song, L. E. Li, and G. Huang, “Vision transformer with deformable attention,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 4794–4803

    2022

  24. [32]

    Ntire 2017 challenge on single image super-resolution: Dataset and study,

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 2017, pp. 1122–1131

    2017

  25. [33]

    Enhanced deep residual networks for single image super-resolution,

    B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee, “Enhanced deep residual networks for single image super-resolution,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , July 2017

    2017

  26. [34]

    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

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.