Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Image Demoir\'eing Using Dual Camera Fusion on Mobile Phones

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Using the ultra-wide-angle photo from the same phone as a reference, the paper removes severe screen moiré from the wide-angle photo and beats previous single-image methods on a new 8,959-sample real-world benchmark.

desk verdict A solid dual-camera demoiréing dataset and method whose main weakness is evaluation breadth, not the core idea. read the letter →

arxiv 2506.08361 v1 pith:S4PF3I5B submitted 2025-06-10 cs.CV

Yanting Mei , Zhilu Zhang , Xiaohe Wu , Wangmeng Zuo This is my paper
classification cs.CV
keywords imagedemoiréingdualcamerafusionultra-wide-anglemobilephotographykeypointmatchingalignmentkernelpredictionreal-worlddatasetscreenmoiréremoval
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

This paper tries to establish that the second camera already on most smartphones—the ultra-wide-angle (UW) lens—can be used to remove moiré from the wide-angle (W) photo of a screen, and that this dual-camera strategy beats single-image demoiréing methods on severe moiré. The authors build a real-world dataset of about 9,000 W/UW/ground-truth triples from three phones and three monitors, and design a network that aligns the UW image to the W image in two stages and fuses their features. On every phone subset in the dataset, the proposed method improves PSNR, SSIM, LPIPS, and ΔE over the best prior single-image method. If the result holds, screen photography on ordinary dual-camera phones can be improved without a large, slow network.

What carries the argument

The machinery is a two-stage alignment between the wide-angle and ultra-wide-angle views. In the first stage, KMA (Keypoint Matching based Alignment) detects 256 keypoints with SuperPoint on downsampled images, matches them with LightGlue, and warps the UW image coarsely onto the W image. In the second stage, KPA (Kernel Prediction based Alignment) pools and concatenates the two encoders' multi-scale features, runs two convolutions, applies a softmax over weight groups, and combines learnable parameters to produce a content-adaptive convolution kernel that refines $\tilde{F}_{uw}$ from the aligned UW features; the fused feature is $F_m + \alpha \tilde{F}_{uw}$ with a learned per-channel coefficient. The same recipe is used to build the dataset: GlueStick keypoint/keyline matching followed by FlowFormer optical flow aligns the source screen image to the cropped W image, and a least-squares 3x3 color correction matrix computed on Gaussian-blurred images maps the source into the ground-truth color state. The alignment is the load-bearing part: it is what converts the UW image from a misaligned second view into usable moiré-free evidence, and the ablation shows that KMA+KPA gives the gain while optical flow or deformable convolution does not.

What would settle it

Measure moiré energy in the UW images of the released DCID dataset: if a substantial share of UW images show moiré comparable to their paired W images, the premise fails. A direct experiment would be to capture W/UW pairs on other phone models and screens at distances and angles where both sensors alias, then run the released model and check whether its PSNR advantage over single-image ESDNet disappears or turns negative.

Watch

Extended reading notes

Core claim

The central claim is that when severe frequency aliasing puts moiré on the wide-angle (W) sensor, the ultra-wide-angle (UW) sensor—having a different focal length and viewpoint—usually records the same screen with normal colors and textures, and that this UW image is a usable reference for removing moiré from the W image. To test this, the paper contributes the DCID dataset, about 8,959 real captures across three phone models and three monitors, with ground truth aligned to the cropped W image and color-corrected to the display's actual appearance. The method wraps the ESDNet demoiréing backbone with a lightweight UW encoder and aligns the two views first by keypoint matching (KMA) at the image level and then by predicted convolution kernels (KPA) at the feature level, fusing the features as $F_{fuse} = F_m + \alpha \tilde{F}_{uw}$. On the Xiaomi subset the paper reports PSNR 27.06 dB versus 26.14 dB for ESDNet, with similar gains on the Huawei and iPhone subsets and lower color error in every case. The authors conclude that the UW image supplies low-resolution but moiré-free texture and color information that a single degraded image cannot provide.

Load-bearing premise

The approach stands or falls on the assumption, reported as an observation rather than measured, that the ultra-wide-angle image is essentially moiré-free and color-faithful whenever the wide-angle image has severe moiré; if the UW camera also aliases with the screen under common conditions, or the two views cannot be aligned, the fusion has no reliable reference to add.

Editorial extensions

If this is right

  • Screen photos taken on dual-camera phones can be demoiréd with better PSNR, SSIM, LPIPS, and ΔE than any single-image method compared, on all three phone subsets of the DCID dataset.
  • Textures hidden under severe moiré—the paper's lamp-shade example—are recoverable from the UW view instead of being hallucinated by the network.
  • The KMA+KPA alignment is what delivers the gain: replacing either stage with optical flow or deformable convolution lowers performance, and KMA+KPA adds only about 0.5 GFLOPs and 0.14 seconds over the ESDNet baseline.
  • Perceptual loss weight $\lambda_p = 2.0$ is needed; without it the best PSNR drops by about 1 dB in the ablation.
  • The released DCID dataset of 8,959 real samples gives future work a benchmark where severe, color-rich moiré is the target rather than light synthetic patterns.

Reading between the lines

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

  • If UW moiré-freeness holds more broadly than the three phones tested, the same architecture could be applied to live screen capture or video, since the first alignment stage runs on downsampled images and the total inference time is under a second.
  • The two-stage alignment is generic: the same KMA+KPA recipe could carry a clean reference view into other single-image restoration tasks where a second lens is available, such as glare removal or reflection suppression.
  • The paper does not quantify how often or how strongly the UW image itself contains moiré; a direct test would be to measure moiré energy in UW images across distances, angles, and screen subpixel layouts, since the whole gain depends on that premise.
  • Adapting existing methods to accept the UW image (the RRID and UnDeM variants in Table III) helps less than the purpose-built fusion, suggesting the alignment and fusion design, not merely the extra input, drives the improvement.
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 DCID (Dual Camera fusion for Image Demoiréing), a method that uses the ultra-wide-angle (UW) image from a smartphone's secondary camera to assist the demoiréing of the wide-angle (W) image. The method integrates a lightweight UW encoder into the ESDNet backbone and introduces a two-stage alignment procedure: KMA (keypoint matching via SuperPoint and LightGlue) for coarse image-level alignment and KPA (content-adaptive kernel prediction) for fine feature-level alignment. The authors also construct a new dataset of approximately 9,000 real-world screen captures using three phones and three monitors, with ground truth generated by aligning the source image to the W image. Experiments on this dataset show consistent improvements over several single-image demoiréing methods, e.g., PSNR 27.06 vs. 26.14 dB for ESDNet on the Xiaomi subset, and ablations indicate that the two-stage alignment is responsible for the gains.

Significance. If the reported results hold, this is the first dual-camera fusion approach for image demoiréing on mobile phones and it establishes a new benchmark dataset (DCID) that the authors plan to release. The work is practically motivated, and the efficiency-oriented design (lightweight UW encoder, fast alignment) is well aligned with mobile deployment. The manuscript includes a relatively thorough ablation of alignment alternatives (optical flow, deformable convolution, KMA, KPA) and a sensitivity study for the perceptual loss weight. The dataset, code release, and the identification of a novel problem setting are valuable contributions. However, the evaluation is confined to the authors' own dataset, the improvements over the strongest baseline are modest, and the analysis lacks quantification of alignment accuracy and run-to-run variance.

major comments (4)
  1. [Section V.C, Table IV] The first row of Table IV (no alignment, no UW) reports PSNR 26.14, which is identical to the ESDNet baseline in Table II. This means the entire reported improvement (0.92 dB on the Xiaomi subset) comes from the KMA+KPA alignment stages, not from the presence of the UW image per se. The paper never measures alignment accuracy directly (e.g., keypoint inlier ratio, reprojection error, or SSIM between aligned UW and W images) and reports no failure cases. Given that the central claim is that UW-assisted fusion beats single-image demoiréing, the absence of any alignment-quality metric is a load-bearing gap. The authors should provide quantitative alignment accuracy statistics and a qualitative analysis of failure modes (e.g., low-texture screens, large parallax, or strong perspective differences).
  2. [Section III.B and Section V] The ground-truth generation pipeline in Section III.B uses GlueStick and FlowFormer to align the source image to the W image, while the proposed KMA uses SuperPoint and LightGlue for keypoint matching. This is not circular because the ground truth is derived from the source image, not from the UW image. However, the test set is drawn from the same three phones and three monitors used to build the dataset, and no evaluation is performed on existing benchmarks such as LCDMoiré, FHDMi, or UHDM. The paper claims state-of-the-art performance 'on our dataset,' but without cross-dataset or external validation, the generality of the method remains unsubstantiated. At minimum, the authors should report the performance of the ESDNet backbone on a standard benchmark to confirm that their training configuration is consistent with published numbers, and ideally discuss how the method would transfer to unseen devices or screen types.
  3. [Section V.C, Table IV] The ablation in Table IV does not include a trivial baseline in which the UW image is naively resized or cropped and concatenated or added to the W branch without any learned geometric alignment. The fact that the no-alignment row matches the ESDNet baseline exactly suggests that the UW encoder and adaptive fusion alone contribute nothing without alignment. A naive-baseline experiment would clarify whether the improvement is specifically attributable to the proposed KMA+KPA alignment or simply to the injection of additional information at the correct spatial location. The authors should add such a baseline (e.g., resize UW to W resolution and concatenate features) to support their attribution of the gains to the two-stage alignment.
  4. [Section V.B, Table II] The quantitative results in Table II are reported as point estimates from a single training run. The improvements over ESDNet are small (0.41 dB on HUAWEI, 0.63 dB on iPhone, 0.92 dB on Xiaomi), and no standard deviation, confidence interval, or significance test is provided. The text in Section V.B states that the method 'significantly outperforms' existing methods, but without repeated runs or statistical testing, this claim is unsupported. The authors should train the main models multiple times (e.g., at least three runs with different seeds) and report mean and standard deviation for all metrics, or at least for the primary PSNR comparison.
minor comments (5)
  1. [Section IV.A] In Section IV.A, ESDNet is cited as reference [19], but the correct reference is [6]; reference [19] is a different paper (Residual Dense Network for image super-resolution). The same citation error appears twice in that subsection.
  2. [Section IV.C, Eq. (2) and (6)] The notation is confusing: in Eq. (2), the symbol W' is used for the concatenated pooled features, but in Eq. (3)-(4), W denotes the kernel weights. Additionally, in Eq. (6), 'Ff use' should be typeset as F_fuse. Please clarify the notation to avoid ambiguity.
  3. [Section III.A] The dataset is described as 'diverse mobile phones and monitors,' but it contains only three phone models and three monitor models. The term 'diverse' is an overstatement; consider describing the dataset as containing three devices per category or adding more devices to support the claim.
  4. [Section V.B, Table III] The adaptations of RRID and UnDeM are described only briefly. For reproducibility, please specify the exact changes made: how the RGB input is provided to RRID (which expects RAW), how pseudo-moiré pairs are generated for UnDeM, and what training data is used for these adapted baselines.
  5. [Section V.C] The ablation study is conducted only on the Xiaomi camera subset. Since the improvements vary across subsets in Table II, please clarify whether the conclusions of the ablation (e.g., KMA+KPA being optimal) also hold for the HUAWEI and iPhone subsets, or add such ablations if feasible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central comparison is an empirical evaluation on a held-out split of a newly collected dataset, with ground truth generated by external alignment tools independent of the proposed model.

full rationale

The paper's central claim—that UW-assisted fusion outperforms single-image demoiréing on the DCID dataset—rests on a held-out train/test split and externally generated ground truth, not on fitted values or self-citation. Ground truth is created by aligning the displayed source image to the captured W image using GlueStick and FlowFormer and then applying a 3x3 least-squares color correction; these operations are independent of the proposed network and are applied identically to all compared methods. The compared baselines are retrained on the same dataset with the same configuration, so the reported PSNR/SSIM differences measure actual restoration behavior rather than construction artifacts. Ablation Table IV does show that fusion without alignment equals the ESDNet baseline, but this is an empirical ablation result, not a circular reduction; it simply indicates that the contribution of the UW input is realized through the alignment stages. The paper contains some self-citations (Refs. 11–15) and a reference-list inconsistency in Section IV-A (citing RDN [19] where ESDNet [6] is meant), but none of these is load-bearing for the main result. The unquantified premise that UW images are mostly moiré-free is an external-validity concern, not a circularity.

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

The central claim rests on the observation about UW moiré-free content and on external alignment tools for GT; no new physical entities are proposed. The main hand-set hyperparameters are lambda_p, keypoint count, and downsampling factor.

free parameters (4)
  • perceptual loss weight lambda_p = 2.0
    Chosen by the ablation in Table V; the central result uses 2.0.
  • keypoint count for KMA = 256
    Selected through 'extensive experiments' in Section IV-C to balance speed and alignment accuracy.
  • downsampling factor for keypoint detection = x4
    Keypoints are detected on x4 downsampled images in KMA (Section IV-C).
  • KPA kernel size K and group count G
    K and G appear in Eq. (3)-(4) but are not given numerical values in the paper, leaving the architecture underspecified.
assumptions (4)
  • domain assumption The UW image generally has less moiré than the W image due to different focal lengths and positions.
    Stated as an observation in the Introduction (Section I) and used to justify the entire fusion approach; it is not quantitatively verified.
  • domain assumption ESDNet is a valid and strong backbone for demoiréing, and its architecture as described in the cited reference can be reproduced.
    The method builds on ESDNet as backbone (Section IV-A), but the paper cites the wrong reference [19] for ESDNet, making the exact backbone specification uncertain.
  • domain assumption Ground truth generated by Gluestick and FlowFormer alignment of the source image is sufficiently accurate for training and evaluation.
    Dataset processing in Section III-B uses these tools; imperfect alignment would bias both training and the reported metrics, though it would affect all compared methods equally.
  • domain assumption The linear 3x3 color correction matrix estimated from Gaussian-blurred images restores display color appearance without removing moiré-related color variation.
    Section III-B; this assumption underlies the GT color fidelity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Image Demoir\'eing Using Dual Camera Fusion on Mobile Phones." pith.science (2026). https://pith.science/paper/S4PF3I5B

@misc{pith2026250608361,
  author       = {Pith},
  title        = {Pith review of: Image Demoir\'eing Using Dual Camera Fusion on Mobile Phones},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4PF3I5B}},
  note         = {Machine review of arXiv:2506.08361}
}
read the original abstract

When shooting electronic screens, moir\'e patterns usually appear in captured images, which seriously affects the image quality. Existing image demoir\'eing methods face great challenges in removing large and heavy moir\'e. To address the issue, we propose to utilize Dual Camera fusion for Image Demoir\'eing (DCID), \ie, using the ultra-wide-angle (UW) image to assist the moir\'e removal of wide-angle (W) image. This is inspired by two motivations: (1) the two lenses are commonly equipped with modern smartphones, (2) the UW image generally can provide normal colors and textures when moir\'e exists in the W image mainly due to their different focal lengths. In particular, we propose an efficient DCID method, where a lightweight UW image encoder is integrated into an existing demoir\'eing network and a fast two-stage image alignment manner is present. Moreover, we construct a large-scale real-world dataset with diverse mobile phones and monitors, containing about 9,000 samples. Experiments on the dataset show our method performs better than state-of-the-art methods. Code and dataset are available at https://github.com/Mrduckk/DCID.

Figures

Figures reproduced from arXiv: 2506.08361 by the authors.

Figure 1
Figure 1. Comparison of image demoireing methods. The sub-figure (a) shows a ´ classic image demoireing with a single input. The sub-figure (b) shows image ´ demoireing using asymmetric camera fusion. ´ moire patterns, they do not have satisfactory performance ´ on severe moire ones, as shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Data processing pipeline in our DCID dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Pipeline of the proposed DCID method. coarse geometric alignment on the image level, and the second performs fine-grained alignment on the feature level. KMA. There is a significant difference in the field of view between UW and W images. In this case, the common optical flow based alignment methods are not cost-effective, i.e., achieving a good alignment performance may come at a high cost. Instead, we propose to p… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of demoireing results among state-of-the-art methods on our dataset. Please zoom in for a better view. ´ TABLE II QUANTITATIVE COMPARISONS BETWEEN OUR MODEL AND STATE-OF-THE-ART METHODS ON OUR DATASETS. ↑ DENOTES THE LARGER THE BETTER, AND ↓ DENOTES T…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages

  1. [19]

    Residual dense network for image super-resolution,

    Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu, “Residual dense network for image super-resolution,” in CVPR, 2018

    work page 2018

  2. [6]

    Towards efficient and scale-robust ultra-high-definition image demoir´eing,

    Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Jiajun Shen, Jia Li, and Xiaojuan Qi, “Towards efficient and scale-robust ultra-high-definition image demoir´eing,” in ECCV, 2022

    work page 2022

  3. [1]

    Moir ´e pattern removal from texture images via low-rank and sparse matrix decomposition,

    Fanglei Liu, Jingyu Yang, and Huanjing Yue, “Moir ´e pattern removal from texture images via low-rank and sparse matrix decomposition,” in IEEE VCIP , 2015

    work page 2015

  4. [2]

    Textured image demoir ´eing via signal decomposition and guided filtering,

    Jingyu Yang, Fanglei Liu, Huanjing Yue, Xiaomei Fu, Chunping Hou, and Feng Wu, “Textured image demoir ´eing via signal decomposition and guided filtering,” IEEE TIP , 2017

    work page 2017

  5. [3]

    Mop moire patterns using mopnet,

    Bin He, Ce Wang, Boxin Shi, and Ling-Yu Duan, “Mop moire patterns using mopnet,” in ICCV, 2019

    work page 2019

  6. [4]

    Image demoireing with learnable bandpass filters,

    Bolun Zheng, Shanxin Yuan, Gregory Slabaugh, and Ales Leonardis, “Image demoireing with learnable bandpass filters,” in CVPR, 2020

    work page 2020

  7. [5]

    Fhde 2 net: Full high definition demoireing network,

    Bin He, Ce Wang, Boxin Shi, and Ling-Yu Duan, “Fhde 2 net: Full high definition demoireing network,” in ECCV, 2020

    work page 2020

  8. [7]

    Flowformer: A transformer architecture for optical flow,

    Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li, “Flowformer: A transformer architecture for optical flow,” in ECCV, 2022

    work page 2022

Show all 31 references
  1. [8]

    Deformable convolutional networks,

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei, “Deformable convolutional networks,” in ICCV, 2017

    2017

  2. [9]

    Moir ´e photo restoration using multiresolution convolutional neural networks,

    Yujing Sun, Yizhou Yu, and Wenping Wang, “Moir ´e photo restoration using multiresolution convolutional neural networks,” IEEE TIP , 2018

    2018

  3. [10]

    Multiscale coarse-to- fine guided screenshot demoir ´eing,

    Duong Hai Nguyen, Se-Ho Lee, and Chul Lee, “Multiscale coarse-to- fine guided screenshot demoir ´eing,” SPL, 2023

    2023

  4. [11]

    Self-supervised learning for real-world super-resolution from dual zoomed observations,

    Zhilu Zhang, Ruohao Wang, Hongzhi Zhang, Yunjin Chen, and Wang- meng Zuo, “Self-supervised learning for real-world super-resolution from dual zoomed observations,” in ECCV, 2022

    2022

  5. [12]

    Self-supervised learning for real-world super-resolution from dual and multiple zoomed observations,

    Zhilu Zhang, Ruohao Wang, Hongzhi Zhang, and Wangmeng Zuo, “Self-supervised learning for real-world super-resolution from dual and multiple zoomed observations,” IEEE TPAMI, 2024

    2024

  6. [13]

    Robust reference- based super-resolution with similarity-aware deformable convolution,

    Gyumin Shim, Jinsun Park, and In So Kweon, “Robust reference- based super-resolution with similarity-aware deformable convolution,” in CVPR, 2020

    2020

  7. [14]

    Self-supervised image restoration with blurry and noisy pairs,

    Zhilu Zhang, RongJian Xu, Ming Liu, Zifei Yan, and Wangmeng Zuo, “Self-supervised image restoration with blurry and noisy pairs,” NeurIPS, 2022

    2022

  8. [15]

    Bracketing is all you need: Unifying image restoration and enhancement tasks with multi-exposure images,

    Zhilu Zhang, Shuohao Zhang, Renlong Wu, Zifei Yan, and Wangmeng Zuo, “Bracketing is all you need: Unifying image restoration and enhancement tasks with multi-exposure images,” ICLR, 2025

    2025

  9. [16]

    Self-adaptively learning to demoir ´e from focused and defocused image pairs,

    Lin Liu, Shanxin Yuan, Jianzhuang Liu, Liping Bao, Gregory Slabaugh, and Qi Tian, “Self-adaptively learning to demoir ´e from focused and defocused image pairs,” NeurIPS, 2020

    2020

  10. [17]

    Aim 2019 challenge on image demoireing: Dataset and study,

    Shanxin Yuan, Radu Timofte, Gregory Slabaugh, and Ale ˇs Leonardis, “Aim 2019 challenge on image demoireing: Dataset and study,” in ICCV Workshops, 2019

    2019

  11. [18]

    Gluestick: Robust image matching by sticking points and lines together,

    R ´emi Pautrat, Iago Su ´arez, Yifan Yu, Marc Pollefeys, and Viktor Larsson, “Gluestick: Robust image matching by sticking points and lines together,” in ICCV, 2023

    2023

  12. [20]

    Super- point: Self-supervised interest point detection and description,

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich, “Super- point: Self-supervised interest point detection and description,” in CVPR Workshops, 2018

    2018

  13. [21]

    Light- glue: Local feature matching at light speed,

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys, “Light- glue: Local feature matching at light speed,” in ICCV, 2023

    2023

  14. [22]

    Transxnet: learning both global and local dynamics with a dual dynamic token mixer for visual recognition,

    Meng Lou, Hong-Yu Zhou, Sibei Yang, and Yizhou Yu, “Transxnet: learning both global and local dynamics with a dual dynamic token mixer for visual recognition,” IEEE TNNLS , 2025

    2025

  15. [23]

    Very deep convolutional networks for large-scale image recognition,

    K Simonyan and A Zisserman, “Very deep convolutional networks for large-scale image recognition,” in ICLR, 2015

    2015

  16. [24]

    Wavelet-based dual-branch network for image demoir ´eing,

    Lin Liu, Jianzhuang Liu, Shanxin Yuan, Gregory Slabaugh, Ale ˇs Leonardis, Wengang Zhou, and Qi Tian, “Wavelet-based dual-branch network for image demoir ´eing,” in ECCV, 2020

    2020

  17. [25]

    Image quality assessment: from error visibility to structural similarity,

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE TIP , 2004

    2004

  18. [26]

    The unreasonable effectiveness of deep features as a perceptual metric,

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018

    2018

  19. [27]

    Real- time image demoireing on mobile devices,

    Yuxin Zhang, Mingbao Lin, Xunchao Li, Han Liu, Guozhi Wang, Fei Chao, Shuai Ren, Yafei Wen, Xiaoxin Chen, and Rongrong Ji, “Real- time image demoireing on mobile devices,” ICLR, 2023

    2023

  20. [28]

    Sgdr: Stochastic gradient descent with warm restarts,

    Ilya Loshchilov and Frank Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” ICLR, 2017

    2017

  21. [29]

    Adam: A method for stochastic optimization,

    Diederik P Kingma, “Adam: A method for stochastic optimization,” ICLR, 2015

    2015

  22. [30]

    Image demoireing in raw and srgb domains,

    Shuning Xu, Binbin Song, Xiangyu Chen, Xina Liu, and Jiantao Zhou, “Image demoireing in raw and srgb domains,” in ECCV, 2024

    2024

  23. [31]

    Learning image demoir ´eing from unpaired real data,

    Yunshan Zhong, Yuyao Zhou, Yuxin Zhang, Fei Chao, and Rongrong Ji, “Learning image demoir ´eing from unpaired real data,” in AAAI, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.