arxiv: 2604.04484 · v1 · submitted 2026-04-06 · 📡 eess.IV · cs.CV

Recognition: 2 theorem links

· Lean Theorem

TM-BSN: Triangular-Masked Blind-Spot Network for Real-World Self-Supervised Image Denoising

Junyoung Park , Youngjin Oh , Nam Ik Cho

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:49 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords self-supervised denoisingblind-spot networkreal-world noisesRGB imagesdemosaicingtriangular maskimage restoration

0 comments

The pith

Triangular-masked convolutions let blind-spot networks handle the diamond-shaped noise correlation in real camera images.

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

Standard blind-spot networks for self-supervised denoising assume that noise at each pixel is independent, but this breaks down in real sRGB photographs because the camera's demosaicing step creates spatial correlations. The paper shows that these correlations form a diamond-shaped pattern with decaying strength away from the center pixel. TM-BSN introduces a triangular mask inside the convolution kernels that excludes exactly the correlated neighbors while still using every other surrounding pixel at full resolution. This removes the need for downsampling, which otherwise changes the noise statistics and discards context. The resulting estimates are distilled into a lightweight U-Net, producing higher accuracy than prior self-supervised methods on real-world benchmarks.

Core claim

The TM-BSN architecture restricts each convolution kernel to its upper-triangular region, producing a diamond-shaped blind spot at the native image resolution. This geometry excludes pixels whose values are linearly dependent on the target through demosaicing weights while retaining all uncorrelated context, allowing the network to learn a clean-signal estimator directly from noisy input without ground-truth images or any resolution reduction.

What carries the argument

Triangular-masked convolution that limits the receptive field to the upper-triangular kernel region, thereby aligning the blind spot with the diamond-shaped spatial correlation induced by demosaicing.

If this is right

The network achieves state-of-the-art results among self-supervised methods on established real-world denoising benchmarks.
Full-resolution processing is maintained without the noise-statistic distortion introduced by downsampling.
Knowledge distillation from multiple blind-spot predictions produces a compact U-Net that runs faster while preserving accuracy.
No separate post-processing step is required after the initial blind-spot predictions.

Where Pith is reading between the lines

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

The same triangular masking principle could be adapted to other camera pipelines whose correlation patterns are known from their ISP stages.
Extending the mask to video frames might allow self-supervised denoising when temporal correlations follow similar geometric rules.
Empirical checks on images from additional sensor types would test whether the diamond pattern holds across different demosaicing algorithms.

Load-bearing premise

The spatial correlation between a pixel and its neighbors in real sRGB images follows a fixed diamond-shaped pattern that a single triangular mask can perfectly separate from useful context at full resolution.

What would settle it

Direct measurement of pixel-wise noise correlations in the raw or sRGB data from multiple cameras, showing that substantial correlation remains between the target pixel and locations outside the triangular mask.

Figures

Figures reproduced from arXiv: 2604.04484 by Junyoung Park, Nam Ik Cho, Youngjin Oh.

**Figure 2.** Figure 2: (a) The demosaicing filter assigns higher weights to spa [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Illustration of receptive field expansion and blind-spot formation using the proposed triangular-masked convolution. The receptive [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Overview of the proposed Triangular-Masked Blind-Spot Network (TM-BSN) architecture. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: Qualitative comparison on SIDD Validation dataset [ [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Qualitative comparison on DND Benchmark dataset [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 8.** Figure 8: Qualitative comparison of denoising results with differ [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

read the original abstract

Blind-spot networks (BSNs) enable self-supervised image denoising by preventing access to the target pixel, allowing clean signal estimation without ground-truth supervision. However, this approach assumes pixel-wise noise independence, which is violated in real-world sRGB images due to spatially correlated noise from the camera's image signal processing (ISP) pipeline. While several methods employ downsampling to decorrelate noise, they alter noise statistics and limit the network's ability to utilize full contextual information. In this paper, we propose the Triangular-Masked Blind-Spot Network (TM-BSN), a novel blind-spot architecture that accurately models the spatial correlation of real sRGB noise. This correlation originates from demosaicing, where each pixel is reconstructed from neighboring samples with spatially decaying weights, resulting in a diamond-shaped pattern. To align the receptive field with this geometry, we introduce a triangular-masked convolution that restricts the kernel to its upper-triangular region, creating a diamond-shaped blind spot at the original resolution. This design excludes correlated pixels while fully leveraging uncorrelated context, eliminating the need for downsampling or post-processing. Furthermore, we use knowledge distillation to transfer complementary knowledge from multiple blind-spot predictions into a lightweight U-Net, improving both accuracy and efficiency. Extensive experiments on real-world benchmarks demonstrate that our method achieves state-of-the-art performance, significantly outperforming existing self-supervised approaches. Our code is available at https://github.com/parkjun210/TM-BSN.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TM-BSN's triangular mask gives a direct geometric way to handle demosaicing correlations at full resolution without downsampling, but the performance edge still needs tighter validation.

read the letter

The paper's main contribution is the triangular-masked convolution that produces a diamond-shaped blind spot aligned with demosaicing-induced correlations in real sRGB images. This lets TM-BSN perform self-supervised denoising at native resolution without the downsampling step that previous BSN methods use. It does well by grounding the architecture in the known geometry of the ISP pipeline rather than learning a mask from data. The addition of knowledge distillation to a lightweight U-Net improves both performance and speed. Releasing the code supports checking the details. The soft spot is the lack of direct evidence that the upper-triangular mask exactly excludes the correlated pixels while keeping all useful context. Real camera pipelines include more than just demosaicing, so the actual correlation pattern might differ in shape or extent. If the experiments do not include measurements of noise correlations or targeted ablations on the mask, the reported gains over existing methods could stem from other design elements. The claim of state-of-the-art results on real-world benchmarks needs those checks to hold up. This work is for people focused on self-supervised denoising for practical camera images where full resolution matters. A reader interested in geometrically informed architectures would get something from it. I think it deserves a serious referee. The core idea is clear and the implementation is available, so peer review can sort out the experimental questions.

Referee Report

2 major / 2 minor

Summary. The paper proposes the Triangular-Masked Blind-Spot Network (TM-BSN) for self-supervised denoising of real-world sRGB images. It identifies that standard blind-spot networks assume pixel-wise noise independence, which fails for real sRGB data due to spatially correlated noise induced by the ISP pipeline (especially demosaicing). To address this, TM-BSN introduces a triangular-masked convolution that restricts the kernel to its upper-triangular region, producing a diamond-shaped blind spot at full resolution that excludes correlated pixels while retaining full context; this eliminates the need for downsampling. The method further employs knowledge distillation from multiple blind-spot predictions into a lightweight U-Net. Extensive experiments on real-world benchmarks are reported to show state-of-the-art performance over existing self-supervised approaches.

Significance. If the geometric mask correctly aligns with the actual noise correlation structure, the approach would provide a principled way to perform blind-spot denoising at native resolution without altering noise statistics or discarding context, potentially improving both accuracy and efficiency for practical camera images. The public code release supports reproducibility and enables direct follow-up work.

major comments (2)

[§3] §3 (Method, triangular-masked convolution): The central design claim is that the upper-triangular mask produces a diamond-shaped blind spot whose excluded pixels exactly match the spatially decaying correlations from Bayer demosaicing. This assumption is load-bearing for the claimed advantage over downsampling BSNs, yet the manuscript provides no direct empirical validation (e.g., measured noise autocorrelation maps on the evaluation datasets or controlled ablations varying mask orientation and decay rate). If the actual correlation extent or orientation deviates after full ISP processing, the mask may either leak correlated noise or unnecessarily discard useful context.
[§4] §4 (Experiments): The reported SOTA gains and outperformance over prior self-supervised methods rest on quantitative tables whose statistical significance, variance across runs, and sensitivity to hyper-parameters are not quantified. Without these, it is difficult to determine whether the observed improvements are robust or attributable to the triangular mask geometry versus other implementation choices.

minor comments (2)

[Abstract] The abstract and introduction could more explicitly name the real-world benchmarks (e.g., SIDD, DND) and list the exact competing self-supervised baselines to allow immediate assessment of the scope of the SOTA claim.
[§3] Notation for the masked convolution kernel (upper-triangular restriction) should be formalized with an equation or diagram in §3 to avoid ambiguity when readers attempt to re-implement the receptive-field geometry.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments. These have helped us identify opportunities to strengthen the empirical support for our design choices and the robustness of the reported results. We address each major comment below.

read point-by-point responses

Referee: [§3] §3 (Method, triangular-masked convolution): The central design claim is that the upper-triangular mask produces a diamond-shaped blind spot whose excluded pixels exactly match the spatially decaying correlations from Bayer demosaicing. This assumption is load-bearing for the claimed advantage over downsampling BSNs, yet the manuscript provides no direct empirical validation (e.g., measured noise autocorrelation maps on the evaluation datasets or controlled ablations varying mask orientation and decay rate). If the actual correlation extent or orientation deviates after full ISP processing, the mask may either leak correlated noise or unnecessarily discard useful context.

Authors: We agree that direct empirical validation would strengthen the central claim. The triangular mask is derived from the known diamond-shaped correlation pattern produced by standard Bayer demosaicing (bilinear interpolation with spatially decaying weights), which is documented in the ISP literature. To address the concern, the revised manuscript will include: (1) noise autocorrelation maps computed on the SIDD and DND evaluation datasets, empirically confirming the diamond-shaped decaying correlations after full ISP processing; (2) controlled ablations varying mask orientation (upper-triangular vs. lower-triangular or rectangular) and extent (different triangular region sizes), showing that the proposed geometry optimally excludes correlated pixels while retaining context and yields the highest performance. These additions will demonstrate that deviations from the chosen mask lead to either noise leakage or reduced accuracy. revision: yes
Referee: [§4] §4 (Experiments): The reported SOTA gains and outperformance over prior self-supervised methods rest on quantitative tables whose statistical significance, variance across runs, and sensitivity to hyper-parameters are not quantified. Without these, it is difficult to determine whether the observed improvements are robust or attributable to the triangular mask geometry versus other implementation choices.

Authors: We acknowledge that additional statistical quantification would increase confidence in attributing the gains specifically to the triangular masking. While the improvements are consistent across benchmarks, the revised manuscript will report: (1) mean and standard deviation of PSNR/SSIM over multiple independent runs (at least three with different random seeds) for TM-BSN and the main competing self-supervised methods; (2) a sensitivity analysis on key hyperparameters including mask size, distillation temperature, and loss weighting, confirming that performance remains stable and that the gains persist across reasonable ranges. These results will help isolate the contribution of the proposed geometry from other design choices. revision: yes

Circularity Check

0 steps flagged

No significant circularity; design follows from domain knowledge of ISP pipeline

full rationale

The TM-BSN triangular mask is introduced explicitly to match the diamond-shaped correlation pattern stated to arise from demosaicing with decaying weights; this is a geometric design choice grounded in camera pipeline knowledge rather than any fitted parameter, self-referential equation, or self-citation chain. No step in the provided derivation reduces a prediction or uniqueness claim back to its own inputs by construction. Experiments on external benchmarks supply the performance evidence independently of the mask definition itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on one domain assumption about the geometry of real sRGB noise correlation and on the effectiveness of the newly introduced triangular mask; no free parameters or invented physical entities are introduced beyond the network architecture itself.

axioms (1)

domain assumption Real-world sRGB noise exhibits spatially correlated patterns that form a diamond shape due to demosaicing in the camera ISP pipeline.
Explicitly stated in the abstract as the origin of the correlation that standard BSNs fail to handle.

invented entities (1)

Triangular-masked convolution no independent evidence
purpose: Restrict the receptive field to create a diamond-shaped blind spot at original resolution that excludes correlated pixels.
New architectural component introduced by the paper.

pith-pipeline@v0.9.0 · 5568 in / 1327 out tokens · 55827 ms · 2026-05-10T19:49:17.241904+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We introduce a triangular-masked convolution that restricts the kernel to its upper-triangular region, creating a diamond-shaped blind spot at the original resolution... This design excludes correlated pixels while fully leveraging uncorrelated context
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

TM-BSN uses a triangular-masked convolution... knowledge distillation to transfer complementary knowledge from multiple blind-spot predictions into a lightweight U-Net

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

46 extracted references · 2 canonical work pages

[1]

A high-quality denoising dataset for smartphone cameras

Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. InProceedings of the IEEE conference on com- puter vision and pattern recognition, pages 1692–1700,
[2]

Noise2self: Blind denoising by self-supervision

Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. InInternational conference on machine learning, pages 524–533. PMLR, 2019. 3

2019
[3]

A non-local algorithm for image denoising

Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In2005 IEEE computer so- ciety conference on computer vision and pattern recognition (CVPR’05), pages 60–65. Ieee, 2005. 1

2005
[4]

Image blind denoising with generative adversarial net- work based noise modeling

Jingwen Chen, Jiawei Chen, Hongyang Chao, and Ming Yang. Image blind denoising with generative adversarial net- work based noise modeling. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 3155–3164, 2018. 2, 6

2018
[5]

Exploring efficient asymmetric blind-spots for self- supervised denoising in real-world scenarios

Shiyan Chen, Jiyuan Zhang, Zhaofei Yu, and Tiejun Huang. Exploring efficient asymmetric blind-spots for self- supervised denoising in real-world scenarios. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 2814–2823, 2024. 1, 2, 3, 5, 6, 7, 8

2024
[6]

Image denoising by sparse 3-d transform- domain collaborative filtering.IEEE Transactions on image processing, 16(8):2080–2095, 2007

Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform- domain collaborative filtering.IEEE Transactions on image processing, 16(8):2080–2095, 2007. 1, 6

2080
[7]

Real-world mobile image denoising dataset with efficient baselines

Roman Flepp, Andrey Ignatov, Radu Timofte, and Luc Van Gool. Real-world mobile image denoising dataset with efficient baselines. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22368–22377, 2024. 2

2024
[8]

srgb real noise synthesizing with neighboring correlation-aware noise model

Zixuan Fu, Lanqing Guo, and Bihan Wen. srgb real noise synthesizing with neighboring correlation-aware noise model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1683– 1691, 2023. 2

2023
[9]

Weighted nuclear norm minimization with applica- tion to image denoising

Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with applica- tion to image denoising. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 2862–2869, 2014. 1, 6

2014
[10]

Gunturk, Y

B.K. Gunturk, Y . Altunbasak, and R.M. Mersereau. Color plane interpolation using alternating projections.IEEE Transactions on Image Processing, 11(9):997–1013, 2002. 3

2002
[11]

Toward convolutional blind denoising of real pho- tographs

Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real pho- tographs. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712–1722,
[12]

Neighbor2neighbor: Self-supervised de- noising from single noisy images

Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised de- noising from single noisy images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781–14790, 2021. 3

2021
[13]

C2n: Practical generative noise modeling for real-world denoising

Geonwoon Jang, Wooseok Lee, Sanghyun Son, and Ky- oung Mu Lee. C2n: Practical generative noise modeling for real-world denoising. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 2350–2359,
[14]

Puca: patch-unshuffle and chan- nel attention for enhanced self-supervised image denoising

Hyemi Jang, Junsung Park, Dahuin Jung, Jaihyun Lew, Ho Bae, and Sungroh Yoon. Puca: patch-unshuffle and chan- nel attention for enhanced self-supervised image denoising. Advances in Neural Information Processing Systems, 36: 19217–19229, 2023. 2, 3, 6, 7, 8

2023
[15]

Self-supervised image denoising with down- sampled invariance loss and conditional blind-spot network

Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim, and Nam Ik Cho. Self-supervised image denoising with down- sampled invariance loss and conditional blind-spot network. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12196–12205, 2023. 6, 7, 8

2023
[16]

Apr-rd: Complemental two steps for self-supervised real image denoising

Hyunjun Kim and Nam Ik Cho. Apr-rd: Complemental two steps for self-supervised real image denoising. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 4257–4265, 2025. 2, 3, 5, 6, 7

2025
[17]

Kingma and Jimmy Ba

Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5

2015
[18]

Noise2void-learning denoising from single noisy images

Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 2129–2137, 2019. 3

2019
[19]

High-quality self-supervised deep image denoising.Ad- vances in neural information processing systems, 32, 2019

Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising.Ad- vances in neural information processing systems, 32, 2019. 2, 3

2019
[20]

Ap- bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network

Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. Ap- bsn: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17725–17734, 2022. 1, 2, 3, 6, 7, 8

2022
[21]

Noise2Noise: Learning Image Restoration without Clean Data

Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2

work page Pith review arXiv 2018
[22]

Spatially adaptive self- supervised learning for real-world image denoising

Junyi Li, Zhilu Zhang, Xiaoyu Liu, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Spatially adaptive self- supervised learning for real-world image denoising. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9924, 2023. 6, 7, 8

2023
[23]

Rethink- ing transformer-based blind-spot network for self-supervised image denoising

Junyi Li, Zhilu Zhang, and Wangmeng Zuo. Rethink- ing transformer-based blind-spot network for self-supervised image denoising. InProceedings of the AAAI Conference on Artificial Intelligence, pages 4788–4796, 2025. 3, 5, 6, 7, 8

2025
[24]

Understanding the effective receptive field in deep convolu- tional neural networks.Advances in neural information pro- cessing systems, 29, 2016

Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolu- tional neural networks.Advances in neural information pro- cessing systems, 29, 2016. 1

2016
[25]

Noise2noiseflow: Realistic camera noise modeling without clean images

Ali Maleky, Shayan Kousha, Michael S Brown, and Mar- cus A Brubaker. Noise2noiseflow: Realistic camera noise modeling without clean images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 17632–17641, 2022. 2

2022
[26]

High- quality linear interpolation for demosaicing of bayer- patterned color images

Henrique S Malvar, Li-wei He, and Ross Cutler. High- quality linear interpolation for demosaicing of bayer- patterned color images. In2004 IEEE international con- ference on acoustics, speech, and signal processing, pages iii–485. IEEE, 2004. 3

2004
[27]

Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections.Advances in neu- ral information processing systems, 29, 2016

Xiaojiao Mao, Chunhua Shen, and Yu-Bin Yang. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections.Advances in neu- ral information processing systems, 29, 2016. 2

2016
[28]

Cvf-sid: Cyclic multi-variate func- tion for self-supervised image denoising by disentangling noise from image

Reyhaneh Neshatavar, Mohsen Yavartanoo, Sanghyun Son, and Kyoung Mu Lee. Cvf-sid: Cyclic multi-variate func- tion for self-supervised image denoising by disentangling noise from image. InProceedings of the ieee/cvf Conference on Computer Vision and Pattern Recognition, pages 17583– 17591, 2022. 6

2022
[29]

Random sub-samples generation for self- supervised real image denoising

Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self- supervised real image denoising. InProceedings of the IEEE/CVF international conference on computer vision, pages 12150–12159, 2023. 2, 3, 6, 8

2023
[30]

Benchmarking denoising algorithms with real photographs

Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 1586–1595, 2017. 5, 7

2017
[31]

U- net: Convolutional networks for biomedical image segmen- tation

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 2, 4, 5

2015
[32]

Nonlinear total variation based noise removal algorithms.Physica D: nonlinear phenomena, 60(1-4):259–268, 1992

Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms.Physica D: nonlinear phenomena, 60(1-4):259–268, 1992. 1

1992
[33]

Variational deep image restoration.IEEE Transactions on Image Processing, 31: 4363–4376, 2022

Jae Woong Soh and Nam Ik Cho. Variational deep image restoration.IEEE Transactions on Image Processing, 31: 4363–4376, 2022. 4

2022
[34]

Mem- net: A persistent memory network for image restoration

Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Mem- net: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on com- puter vision, pages 4539–4547, 2017. 2

2017
[35]

Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 5

2004
[36]

Lg-bpn: Local and global blind-patch network for self-supervised real-world denoising

Zichun Wang, Ying Fu, Ji Liu, and Yulun Zhang. Lg-bpn: Local and global blind-patch network for self-supervised real-world denoising. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 18156–18165, 2023. 1, 3, 6, 7

2023
[37]

Primary-consistent soft- decision color demosaicking for digital cameras (patent pending).IEEE Transactions on Image Processing, 13(9): 1263–1274, 2004

Xiaolin Wu and Ning Zhang. Primary-consistent soft- decision color demosaicking for digital cameras (patent pending).IEEE Transactions on Image Processing, 13(9): 1263–1274, 2004. 3

2004
[38]

Unpaired learning of deep image denoising

Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wang- meng Zuo. Unpaired learning of deep image denoising. In European conference on computer vision, pages 352–368. Springer, 2020. 2, 3

2020
[39]

Real-world noisy image denoising: A new bench- mark.arXiv preprint arXiv:1804.02603, 2018

Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new bench- mark.arXiv preprint arXiv:1804.02603, 2018. 2

work page arXiv 2018
[40]

Variational denoising network: Toward blind noise modeling and removal.Advances in neural information processing systems, 32, 2019

Zongsheng Yue, Hongwei Yong, Qian Zhao, Deyu Meng, and Lei Zhang. Variational denoising network: Toward blind noise modeling and removal.Advances in neural information processing systems, 32, 2019. 6

2019
[41]

Cycleisp: Real image restoration via improved data synthesis

Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Cycleisp: Real image restoration via improved data synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2696–2705,
[42]

Restormer: Efficient transformer for high-resolution image restoration

Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728–5739,
[43]

Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE transactions on image processing, 26(7):3142–3155, 2017

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. 1, 2, 6

2017
[44]

Ffdnet: Toward a fast and flexible solution for cnn-based image denoising

Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608–4622,
[45]

Color demosaicking via di- rectional linear minimum mean square-error estimation

Lei Zhang and Xiaolin Wu. Color demosaicking via di- rectional linear minimum mean square-error estimation. IEEE Transactions on Image Processing, 14(12):2167–2178,
[46]

When awgn-based denoiser meets real noises

Yuqian Zhou, Jianbo Jiao, Haibin Huang, Yang Wang, Jue Wang, Honghui Shi, and Thomas Huang. When awgn-based denoiser meets real noises. InProceedings of the AAAI Con- ference on Artificial Intelligence, pages 13074–13081, 2020. 3

2020