Pith. sign in

REVIEW 4 major objections 6 minor 51 references

JPEG Processing Neural Operator for Backward-Compatible Coding

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read JPNeO upgrades JPEG with learned encoder and decoder while keeping the stored format fully compatible.

desk verdict A well-engineered backward-compatible JPEG codec with strong results, but the SOTA claim is missing the two closest encoder-side baselines. read the letter →

arxiv 2507.23521 v1 pith:F4U72GGQ submitted 2025-07-31 eess.IV cs.CV

Woo Kyoung Han , Yongjun Lee , Byeonghun Lee , Sang Hyun Park , Sunghoon Im , Kyong Hwan Jin This is my paper
classification eess.IVcs.CV
keywords JPEGcompressionneuraloperatorbackwardcompatibilitychromarestorationartifactremovalquantizationmatrixmutualinformationlearnedcodec
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 JPNeO, a learned image codec that wraps the existing JPEG standard with neural operators on both the encoding and decoding side while leaving the stored bitstream format untouched. The aim is to obtain compression-quality gains that would otherwise require a new standard: files produced by JPNeO's encoder remain decodable by any conventional JPEG decoder, and JPNeO's decoder can read ordinary JPEG files. The authors report that JPNeO outperforms existing artifact-removal networks at low bitrates, especially at quality factor 0, and that the advantage is driven by chroma restoration. They also report substantially lower memory use and parameter counts than the comparison methods. If these claims hold, the practical payoff is a drop-in upgrade path for existing JPEG pipelines, from cameras to image servers.

What carries the argument

The load-bearing machinery is two neural operators that treat the image as a continuous function. JENO uses a feature extractor, a sampling operator, and a Galerkin-attention neural operator to predict the high-frequency content lost during 4:2:0 or 4:2:2 chroma subsampling, effectively acting as a high-pass filter placed before the standard JPEG encoder. JDNO uses group embedding of the DCT spectra, a Swin-based feature extractor, and a Cosine Neural Operator that predicts continuous cosine coefficients and thereby acts as a learned inverse transform. The learned quantization matrix is initialized from the standard quality-50 matrix through one linear layer and stored as a lookup table after training, so it adds no computational cost at deployment.

What would settle it

Take a fixed set of images, compress them with a standard quality-0 quantization matrix, decode once with the trained JDNO and once with standard JPEG decoding plus bilinear chroma upsampling, and compare PSNR and SSIM; if JDNO shows no meaningful gain at q=0, the central low-bitrate claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a fully JPEG-compatible codec can beat both the JPEG baseline and learned artifact-removal decoders by adding learnable mutual information at both ends of the codec. JPNeO pairs JENO, a preprocessing encoder that reduces the information lost in chroma subsampling, with JDNO, a decoder that maps quantized DCT spectra directly to images through a continuous cosine neural operator, and a learned quantization matrix that tunes the rate–distortion trade-off. On the LIVE1 and BSDS500 test sets, the full system reports higher PSNR and SSIM than prior artifact-removal networks for quality factors 0 through 40, with the largest margin at quality factor 0; the module-level analysis says JENO raises the high-bitrate ceiling while JDNO raises the low-bitrate floor. All of this is done while keeping the output decodable by an unmodified JPEG decoder.

Load-bearing premise

The load-bearing premise is that the neural decoder, trained on quality factors 10 through 90 plus learned quantization matrices, continues to work well at quality factor 0, where the headline low-bitrate gains are reported.

Editorial extensions

If this is right

  • Any bitstream produced by the JENO encoder is decodable by an unmodified JPEG decoder, so the encoder side could be deployed in cameras or ISP pipelines without changing readers.
  • The JDNO decoder can be used standalone to clean up legacy JPEG files, including very low quality factors, without re-encoding.
  • Because JENO and JDNO are decoupled, a deployment can choose one or the other depending on bitrate, saving memory while keeping most of the quality gain.
  • The reported per-module gains indicate the same recipe—separate learned pre- and post-processing around a fixed legacy codec—could be applied to other transform-based codecs.

Reading between the lines

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

  • Beyond the paper's experiments, the extrapolation to quality factor 0 could be tested directly: with the decoder trained only on quality factors 10–90, a held-out q=0 evaluation would reveal whether the low-bitrate advantage is genuine generalization or an artifact of the training distribution.
  • The mutual-information analysis suggests a resource-allocation rule the authors state qualitatively: use the neural encoder alone at high bitrates, the neural decoder alone at low bitrates, and both in between; a practical implementation could switch modules based on the target bitrate.
  • Because the reported gains concentrate in chroma, a targeted stress test on color-saturated content—graphics, skin tones, foliage—would indicate whether the chroma restoration transfers beyond natural-image benchmarks.
  • The 17 stored quantization matrices invite the obvious extension: interpolate between them to hit a desired bitrate more finely, which the paper does not explore.
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 / 6 minor

Summary. The paper presents JPNeO, a neural JPEG codec that is backward compatible with the existing JPEG format. It consists of three components: a JPEG Encoding Neural Operator (JENO) that pre-processes images (mainly targeting chroma subsampling loss), a learned quantization matrix Qψ trained in the spirit of Strümpler et al., and a JPEG Decoding Neural Operator (JDNO) that reconstructs images directly from DCT spectra. The authors claim state-of-the-art performance within the JPEG compression framework, especially at very low bitrates and for chroma quality, and validate a mutual-information-based design rationale. Experiments on LIVE1 and BSDS500 compare JPNeO with artifact-removal networks (DnCNN, QGAC, FBCNN, JDEC), and additional ablations measure computational cost and memory. Code is released.

Significance. If the results hold, the contribution is practically relevant: a backward-compatible codec that improves JPEG without changing the bitstream or decoder protocol could be adopted in existing ISP pipelines. The paper has explicit strengths: the architecture is parameter-efficient (29.7M parameters versus 38.9M for JDEC), the module-wise ablations and resource measurements are useful, and the authors provide source code. The main weaknesses are the missing quantitative comparison against the closest encoder-side backward-compatible methods (Strümpler et al., Talebi et al.) and a few empirical-reporting issues (q=0 extrapolation and an apparent data error in Table 2). The mutual-information rationale is suggestive but not rigorously established.

major comments (4)
  1. [§5.3, Tables 1–2] The central claim of state-of-the-art performance within the JPEG compression framework is not supported by the baseline set: Tables 1 and 2 compare only against artifact-removal decoders (DnCNN, QGAC, FBCNN, JDEC), while Strümpler et al. [41] and Talebi et al. [42], which are the closest encoder-side backward-compatible methods and are cited in Sec. 2, are absent from all quantitative comparisons. Since Qψ is explicitly derived from [41], the authors should report rate-distortion results for these methods on the same LIVE1/BSDS500 protocol; without this, the SOTA claim is unsupported at every quality factor.
  2. [§5.2 / §5.3, Table 1] The evaluation includes quality factors q=0 and q=5, but JDNO is trained only on Q from [10,90] and on the learned Qψ matrices; the largest reported gains occur at q=0 (e.g., LIVE1 PSNR 23.15 for JPNeO versus 21.70 for FBCNN). The manuscript does not discuss this extrapolation regime or verify that the learned Qψ covers it. Please provide additional analysis (e.g., a train/validation split by quality factor, or an explicit statement that the q=0 results are a generalization test) so that the low-bitrate claim can be evaluated.
  3. [Table 2] In the BSDS500 q=10 row, the reported chroma SSIM for JPNeO is 0.138, which is well below all baselines (e.g., FBCNN 0.302) and below JPNeO's own q=0 value of 0.164. This is internally inconsistent with the text's claim that improved chroma restoration drives JPNeO's gains. The value appears to be a typographical error, but it must be corrected and the surrounding conclusion re-verified.
  4. [§3, Eqs. (4)–(9), Fig. 13] The mutual-information justification is informal and the empirical validation is not reproducible. In Eq. (4), I(X̃; θ̂) treats the trained parameter vector θ̂ as a random variable jointly distributed with a single test image, which is not a standard interpretation; the decomposition is never derived. Fig. 13 reports mutual information with the ground-truth image but does not describe the estimator, the binning of symbol statistics, or the sample size. The paper should either formalize the quantity being measured or soften the claim that the hypothesis is experimentally validated.
minor comments (6)
  1. [Table 1] The column header and the '#Params' row are misaligned; DnCNN appears in the data rows but is missing from the header, and the parameter count 259.4M is not elsewhere attributed. Please render the table so every column is labeled.
  2. [Eq. (20)] The symbol ⊗ is used for both the Kronecker product and the elementwise modulation of the cosine features; please define the intended broadcasting/operation explicitly.
  3. [§5.1] The statement 'we set λ by sampling the values from [1e-4, 1e6]' is unclear about the sampling distribution and the number of λ values; please specify the exact procedure and how the 17 stored Qψ correspond to the sampled λ values.
  4. [§6 / Fig. 13] The mutual-information curve would benefit from a caption explaining the estimator; in addition, the t-SNE plot in Fig. 14 does not directly quantify mutual information and should be framed as qualitative evidence only.
  5. [§5.3 / Supplementary Tab. S3] The supplementary ablation in Tab. S3 and Fig. S6 supports the neural-operator design, but the main paper does not mention it; consider adding a sentence in Sec. 5.3 pointing to this ablation.
  6. [Figures 3 and 10] There are typographical errors in the text: 'Opearator' in Fig. 3 and 'cu rve' in the Fig. 10 caption should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central RD claims rest on independent artifact-removal baselines; MI argument is a heuristic sanity check, and the missing Strümpler comparison is an empirical-support gap, not a circular step.

full rationale

The main performance claim is established by comparing JPNeO against independently published artifact-removal decoders (DnCNN, QGAC, FBCNN, JDEC) on held-out LIVE1 and BSDS500 under standard quantization Q (Sec. 5.3, Tabs. 1-2, Fig. 10). The encoder and decoder are trained with explicit reconstruction losses (Eqs. 15 and 23), and no fitted parameter is renamed as a prediction in the RD tables; the q=0 evaluation is extrapolation outside the training quality range, which is a generalization risk, not circularity. The mutual-information discussion (Eqs. 4-9, Fig. 13) is heuristic and partly self-confirmatory---the 'increase' in I(X;Xhat) is measured on networks trained to maximize reconstruction fidelity---so it should be read as a sanity check rather than as an independent first-principles prediction. JDEC [22] is a self-citation and is used as architectural inspiration for JDNO, but JDEC is also included as an external baseline and is outperformed on the same protocol, so the citation is not load-bearing circular support. The main empirical-support gap is that the 'state-of-the-art within the JPEG compression framework' claim is not tested against the closest encoder-side JPEG-compatible methods, Strümpler et al. [41] and Talebi et al. [42], which are cited but absent from all comparisons; this is a baseline-completeness/correctness issue, not a circularity defect. Overall, the derivation chain is self-contained and no claimed result reduces by construction to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method's performance is empirical and rests on the JPEG standard plus several informal information-theoretic approximations. The main fitted component is Qψ; no new physical entities are introduced. The network weights of JENO and JDNO are learned but are not listed as free parameters because they are standard trained model parameters.

free parameters (2)
  • Learned quantization matrix Qψ = 8x8 integer matrix, 17 variants
    Trained with loss λ Ld + Lr on DIV2K/Flickr2K; stored as a lookup table and used in encoding. Directly affects the rate-distortion trade-off and is a fitted component, not derived from first principles.
  • Loss trade-off λ = 17 sampled values in [1e-4, 1e6]
    Chosen by hand to control Qψ training; the set of 17 Qψ values and thus the achievable RD performance depends on this sampling.
assumptions (4)
  • standard math Standard JPEG codec operations as described in Eq. (3): chroma subsampling, DCT, quantization, entropy coding
    Foundation of the problem formulation; treated as ground truth for the backward-compatible constraints.
  • standard math Data processing inequality applies so that I(X; X̃) is limited by the JPEG encoder
    Used in Sec. 3 to motivate the neural decoder; standard information theory, though the application is informal.
  • ad hoc to paper Mutual information decomposition I(X; D̂θ(E_JPEG(X))) ≈ I(X; X̃) + I(X̃; θ̂) (Eq. 4)
    Unproved approximation; central to the MI motivation but not a theorem, and the estimation method is not described.
  • ad hoc to paper Trained JENO satisfies U(Eφ(X)) ≈ HPF(X) (Eqs. 16-18)
    Interpretation of the learned operator as a high-pass filter; not a design guarantee, only an empirical explanation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of JPEG Processing Neural Operator for Backward-Compatible Coding." pith.science (2026). https://pith.science/paper/F4U72GGQ

@misc{pith2026250723521,
  author       = {Pith},
  title        = {Pith review of: JPEG Processing Neural Operator for Backward-Compatible Coding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4U72GGQ}},
  note         = {Machine review of arXiv:2507.23521}
}
read the original abstract

Despite significant advances in learning-based lossy compression algorithms, standardizing codecs remains a critical challenge. In this paper, we present the JPEG Processing Neural Operator (JPNeO), a next-generation JPEG algorithm that maintains full backward compatibility with the current JPEG format. Our JPNeO improves chroma component preservation and enhances reconstruction fidelity compared to existing artifact removal methods by incorporating neural operators in both the encoding and decoding stages. JPNeO achieves practical benefits in terms of reduced memory usage and parameter count. We further validate our hypothesis about the existence of a space with high mutual information through empirical evidence. In summary, the JPNeO functions as a high-performance out-of-the-box image compression pipeline without changing source coding's protocol. Our source code is available at https://github.com/WooKyoungHan/JPNeO.

Figures

Figures reproduced from arXiv: 2507.23521 by the authors.

Figure 1
Figure 1. Overall Concept of the proposed JPNeO method. Our method enables flexible switching between the conventional JPEG encoder and decoder as needed. Our approach ensures interoper￾ability, allowing conventional JPEG-encoded files to be decoded with a conventional JPEG decoder and existing JPEG files to be decoded using our method. of the image processing pipeline. Therefore, research on DNN-based approaches for JPEG is … view at source ↗
Figure 2
Figure 2. Illustration of image sets under the JPEG process. Eφ/JPEG and Dθ/JPEG indicate the neural/JPEG encoder and de￾coder, respectively. Eφ maps images to a range with lower distor￾tion than the EJPEG range. DJPEG is limited to its original range. Our JPNeO’s Dθ directs images to a range with lower distortion. are relatively closer to the ground truth. As demonstrated in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Schematic overview of our JPNeO. JPNeO consists of a JPEG encoding neral operator (JENO (Eφ)), a pre-trained quantization matrix (Qψ), and a JPEG decoding neural operator (JDNO (Dθ)). Our JENO, consisting of a feature extractor (fξ), sampling (S), and neural operators (Gϕ) functions as an auxiliary encoder to the existing JPEG encoder (EJPEG). JDNO, conposed of a group embedding (g), feature extractor fξ, and Cosine… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Overview of learning the quantization matrix. Only one linear layer is trained. Since quantization is not differentiable, it is replaced with an approximation operation (blue area). Unlike existing work [41] that uses trainable parameters, we use a standard quantizatio…
Figure 6
Figure 6. Figure 6: Schematic flow of Cosine Neural Operator (CNO). Eq. (13) as below: \mathbf {\hat X} &= U(E_\varphi (\mathbf {X})) + U(\mathbf {X}')\\ \ &\simeq HPF(\mathbf {X}) + LPF(\mathbf {X}) \\ \therefore U(E_\varphi (\mathbf {X})) &\simeq HPF(\mathbf {X}) \quad (\because U(\math…
Figure 7
Figure 7. Figure 7: Quantitative ablation study on our JENO (top), JDNO (bottom), and [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Qualitative ablation study in highly compressed images based on the quantization map ( [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison of chroma components in high-bpp images based on the encoder (JENO) and the quantization matrix [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 10
Figure 10. Figure 10: Rate-Distortion cuvre comparisons with the [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: Qualitative Comparison in color JPEG artifact removal against JPEG artifact removal networks (q = 5 (top), q = 0 (bottom)). q = 10 Chroma DnCNN [48] QGAC [17] FBCNN [23] JDEC [22] JPNeO GT [PITH_FULL_IMAGE:figures/full_fig_p007_11.png]
Figure 12
Figure 12. Figure 12: Qualitative Comparison in color JPEG artifact removal with chroma components (q = 10). Methods JPEG[45] QGAC[17] FBCNN[23] JDEC[22] JPNeO Dataset q #Params. 259.4M 70.1M 38.9M 29.7M LIVE-1 [39] 0 20.89|19.73 16.33|15.99 21.70|21.19 20.76|20.07 23.15|22.64 0.540 0.318 …
Figure 14
Figure 14. Figure 14: t-SNE [44] clustering analysis by quality factor on the LIVE-1 [39]. Encoded images with our JENO (Eφ) (reds) and JPEG encoder (EJPEG) (blues). we calculate the mutual information using the symbol statis￾tics from the Kodak dataset [19] while switching the en￾coder an…
Figure 13
Figure 13. Figure 13: Mutual information with the ground￾truth image (I(X; Xˆ )) across quality levels in Kodak [19] (4:2:0 subsampling). The label indicates ‘Encoder (Eφ/EJPEG)+‘Q/Qψ ’+Decoder (Dθ/DJPEG)’ ond row of [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 47 canonical work pages

  1. [41]

    Learning to improve image compression without changing the stan- dard decoder

    Yannick Str ¨umpler, Ren Yang, and Radu Timofte. Learning to improve image compression without changing the stan- dard decoder. In European Conference on Computer Vision, pages 200–216. Springer, 2020. 1, 2, 3, 4

    work page 2020

  2. [42]

    Better compression with deep pre-editing

    Hossein Talebi, Damien Kelly, Xiyang Luo, Ignacio Garcia Dorado, Feng Yang, Peyman Milanfar, and Michael Elad. Better compression with deep pre-editing. IEEE Transac- tions on Image Processing, 30:6673–6685, 2021. 1, 2

    work page 2021

  3. [1]

    NTIRE 2017 Chal- lenge on Single Image Super-Resolution: Dataset and Study

    Eirikur Agustsson and Radu Timofte. NTIRE 2017 Chal- lenge on Single Image Super-Resolution: Dataset and Study. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5

    work page 2017

  4. [2]

    Natarajan, and Kamisetty R Rao

    Nasir Ahmed, T. Natarajan, and Kamisetty R Rao. Discrete cosine transform. IEEE transactions on Computers, 100(1): 90–93, 1974. 3

    work page 1974

  5. [3]

    Jpeg xl next-generation image compression ar- chitecture and coding tools

    Jyrki Alakuijala, Ruud Van Asseldonk, Sami Boukortt, Mar- tin Bruse, Iulia-Maria Coms,a, Moritz Firsching, Thomas Fis- chbacher, Evgenii Kliuchnikov, Sebastian Gomez, Robert Obryk, et al. Jpeg xl next-generation image compression ar- chitecture and coding tools. In Applications of digital image processing XLII, pages 112–124. SPIE, 2019. 2

    work page 2019

  6. [4]

    Contour detection and hierarchical image seg- mentation

    Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Ji- tendra Malik. Contour detection and hierarchical image seg- mentation. IEEE transactions on pattern analysis and ma- chine intelligence, 33(5):898–916, 2010. 5, 6, 7, 8, 3

    work page 2010

  7. [5]

    White paper on jpeg ai scope and framework

    Jo ´ao Ascenso and Evgeniy Upenik. White paper on jpeg ai scope and framework. ISO/IEC JTC, 1, 2021. 1

    work page 2021

  8. [6]

    Learning-based image coding: early solutions re- viewing and subjective quality evaluation

    Joao Ascenso, Pinar Akyazi, Fernando Pereira, and Touradj Ebrahimi. Learning-based image coding: early solutions re- viewing and subjective quality evaluation. In Optics, Pho- tonics and Digital Technologies for Imaging Applications VI, pages 164–176. SPIE, 2020. 1

    work page 2020

Show all 51 references
  1. [7]

    What’s in the image? ex- plorable decoding of compressed images

    Yuval Bahat and Tomer Michaeli. What’s in the image? ex- plorable decoding of compressed images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2908–2917, 2021. 2

    2021

  2. [8]

    End-to-end optimized image compression

    Johannes Ball ´e, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 1, 2

    2016 arXiv

  3. [9]

    Variational image compression with a scale hyperprior

    Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436 ,

    arXiv

  4. [10]

    Cas-cnn: A deep convolutional neural network for image compression artifact suppression

    Lukas Cavigelli, Pascal Hager, and Luca Benini. Cas-cnn: A deep convolutional neural network for image compression artifact suppression. In 2017 International Joint Conference on Neural Networks (IJCNN) , pages 752–759. IEEE, 2017. 2

    2017

  5. [11]

    Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration

    Yunjin Chen and Thomas Pock. Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 39(6):1256–1272, 2017. 2

    2017

  6. [12]

    Learning Con- tinuous Image Representation With Local Implicit Image Function

    Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning Con- tinuous Image Representation With Local Implicit Image Function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8628–8638, 2021. 2, 4

    2021

  7. [13]

    Swin2sr: Swinv2 transformer for compressed im- age super-resolution and restoration

    Marcos V Conde, Ui-Jin Choi, Maxime Burchi, and Radu Timofte. Swin2sr: Swinv2 transformer for compressed im- age super-resolution and restoration. In European Confer- ence on Computer Vision , pages 669–687. Springer, 2022. 5

    2022

  8. [14]

    Elements of information theory

    Thomas M Cover. Elements of information theory . John Wiley & Sons, 1999. 1, 3

    1999

  9. [15]

    Compression artifacts reduction by a deep convolu- tional network

    Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression artifacts reduction by a deep convolu- tional network. In Proceedings of the IEEE international conference on computer vision, pages 576–584, 2015. 1, 2

    2015

  10. [16]

    Image super-resolution using deep convolutional net- works

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional net- works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295–307, 2015

    2015

  11. [17]

    Quantization guided jpeg artifact correction

    Max Ehrlich, Larry Davis, Ser-Nam Lim, and Abhinav Shri- vastava. Quantization guided jpeg artifact correction. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII 16, pages 293–309. Springer, 2020. 1, 2, 5, 6, 7

    2020

  12. [18]

    Theory and prac- tice of finite elements

    Alexandre Ern and Jean-Luc Guermond. Theory and prac- tice of finite elements. Springer, 2004. 2

    2004

  13. [19]

    Kodak lossless true color image suite

    Rich Franzen. Kodak lossless true color image suite. source: http://r0k. us/graphics/kodak, 4(2), 1999. 8

    1999

  14. [20]

    Practical learned lossless jpeg recompression with multi-level cross-channel entropy model in the dct domain

    Lina Guo, Xinjie Shi, Dailan He, Yuanyuan Wang, Rui Ma, Hongwei Qin, and Yan Wang. Practical learned lossless jpeg recompression with multi-level cross-channel entropy model in the dct domain. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogniti...

    2022

  15. [21]

    Multiwavelet-based operator learning for differential equations

    Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations. Advances in neural information processing systems, 34:24048–24062, 2021. 2

    2021

  16. [22]

    Woo Kyoung Han, S. Im, J. Kim, and Kyong Hwan Jin. JDEC: Jpeg decoding via enhanced continuous cosine coef- ficients. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 4, 5, 6, 7, 8

    2024

  17. [23]

    Towards flex- ible blind jpeg artifacts removal

    Jiaxi Jiang, Kai Zhang, and Radu Timofte. Towards flex- ible blind jpeg artifacts removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4997–5006, 2021. 1, 2, 5, 6, 7, 8, 3

    2021

  18. [24]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5

    2015

  19. [25]

    Neural operator: Learning maps between function spaces with applications to pdes

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Aziz- zadenesheli, Kaushik Bhattacharya, Andrew Stuart, and An- ima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Ma- chine Learning Research, 24(89):1–97, 2023. 2, 4

    2023

  20. [26]

    Local texture estima- tor for implicit representation function

    Jaewon Lee and Kyong Hwan Jin. Local texture estima- tor for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1929–1938, 2022. 2, 4 9

    1929

  21. [27]

    Learn- ing local implicit fourier representation for image warping

    Jaewon Lee, Kwang Pyo Choi, and Kyong Hwan Jin. Learn- ing local implicit fourier representation for image warping. In European Conference on Computer Vision (ECCV), pages 182–200. Springer, 2022. 2

    2022

  22. [28]

    Fourier neural operator for para- metric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for para- metric partial differential equations. arXiv preprint arXiv:2010.08895, 2020. 2

    2010 arXiv

  23. [29]

    SwinIR: Image Restoration Using Swin Transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image Restoration Using Swin Transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, pages 1833–1844, 2021. 5

    2021

  24. [30]

    Enhanced Deep Residual Networks for Single Image Super-Resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced Deep Residual Networks for Single Image Super-Resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5, 3

    2017

  25. [31]

    Multi-level wavelet-cnn for image restora- tion

    Pengju Liu, Hongzhi Zhang, Kai Zhang, Liang Lin, and Wangmeng Zuo. Multi-level wavelet-cnn for image restora- tion. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages 773–782,

  26. [32]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218–229, 2021. 2

    2021

  27. [33]

    Joint autoregressive and hierarchical priors for learned im- age compression

    David Minnen, Johannes Ball ´e, and George D Toderici. Joint autoregressive and hierarchical priors for learned im- age compression. Advances in neural information processing systems, 31, 2018. 1, 2

    2018

  28. [34]

    Image sensors and signal processing for digital still cameras

    Junichi Nakamura. Image sensors and signal processing for digital still cameras. CRC press, 2017. 1

    2017

  29. [35]

    Towards bidi- rectional arbitrary image rescaling: Joint optimization and cycle idempotence

    Zhihong Pan, Baopu Li, Dongliang He, Mingde Yao, Wen- hao Wu, Tianwei Lin, Xin Li, and Errui Ding. Towards bidi- rectional arbitrary image rescaling: Joint optimization and cycle idempotence. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognitio...

    2022

  30. [36]

    Rgb no more: Minimally- decoded jpeg vision transformers

    Jeongsoo Park and Justin Johnson. Rgb no more: Minimally- decoded jpeg vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22334–22346, 2023. 1, 5

    2023

  31. [37]

    Wire: Wavelet implicit neural representations

    Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Bara- niuk. Wire: Wavelet implicit neural representations. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507–18516, 2023. 2

    2023

  32. [38]

    C. E. Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379–423, 1948. 1

    1948

  33. [39]

    A statistical evaluation of recent full reference image quality assessment algorithms

    Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image pro- cessing, 15(11):3440–3451, 2006. 5, 6, 7, 8, 2, 3

    2006

  34. [40]

    Implicit Neural Represen- tations with Periodic Activation Functions

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit Neural Represen- tations with Periodic Activation Functions. In Advances in Neural Information Processing Systems , pages 7462–7473. Curran Associates, Inc., 2020. 2

    2020

  35. [43]

    Jpeg 2000: Image compression fundamentals, standards and practice, 2002

    D Taubman. Jpeg 2000: Image compression fundamentals, standards and practice, 2002. 2

    2000

  36. [44]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research , 9 (11), 2008. 8

    2008

  37. [45]

    G.K. Wallace. The jpeg still picture compression standard. IEEE Transactions on Consumer Electronics , 38(1):xviii– xxxiv, 1992. 1, 2, 3, 7

    1992

  38. [46]

    Super-resolution neural oper- ator

    Min Wei and Xuesong Zhang. Super-resolution neural oper- ator. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 18247–18256,

  39. [47]

    Im- plicit Transformer Network for Screen Content Image Con- tinuous Super-Resolution

    Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Im- plicit Transformer Network for Screen Content Image Con- tinuous Super-Resolution. In Advances in Neural Informa- tion Processing Systems, 2021. 2

    2021

  40. [48]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142–3155, 2017. 6, 7 10 JPEG Processing Neural Operator for Backward-Compatible Coding S...

    2017

  41. [49]

    We provide a notation table in Tab

    Introduction This supplementary material is intended to support the main paper. We provide a notation table in Tab. S1 to clarify the methodology and problem formulation. Additional back- ground of JPEG compression codec is provided to define the problem and illustrate how the...

  42. [50]

    3 of the main paper and elaborate on the fundamentals of JPEG compression

    JPEG Preliminary In this section, we build upon Sec. 3 of the main paper and elaborate on the fundamentals of JPEG compression. We then characterize the compression artifacts of JPEG and provide further details of our proposed method. Standard JPEG compression Fig. S1 hows the...

  43. [51]

    JDNO lifting

    Additional Experiments Memory Consumption Following the Sec. 6 of the main paper, we report memory consumption and the number of parameters. We compare our JPNeO with QGAC [17], FBCNN [23], and JDEC [22]. The size of the input for the comparison is 560 × 560. Fig. S3 shows the...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.