Pith. sign in

REVIEW 2 major objections 6 minor 68 references

Towards Better De-raining Generalization via Rainy Characteristics Memorization and Replay

T0 review · 2 major / 6 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 Rain replay memory lets de-raining nets beat prior continual methods

desk verdict A serious continual-learning-for-de-raining paper with broad experiments, but the SPA-data generalization claim is not a clean holdout because λ was chosen on SPA-data. read the letter →

arxiv 2506.02477 v1 pith:FWEKCIO4 submitted 2025-06-03 cs.CV

Kunyu Wang , Xueyang Fu , Chengzhi Cao , Chengjie Ge , Wei Zhai , Zheng-Jun Zha This is my paper
classification cs.CV
keywords imagede-rainingcontinuallearninggenerativereplaycatastrophicforgettingknowledgedistillationgeneralizationcomplementarysystemrainstreakgeneration
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 claims that a de-raining network can keep learning from a growing stream of rainy-image datasets without forgetting earlier ones, and that doing so makes it generalize better to real-world rain it has never seen. The proposed framework, CLGID, borrows the brain's complementary learning system: a GAN per dataset stores the rain-streak characteristics (the hippocampus), and during training of each new dataset the network sees a mix of new images and replayed images built by adding remembered rain layers to clean backgrounds, plus a consistency loss that keeps the network's current outputs aligned with its previous ones. On three different de-raining backbones and streams of up to six datasets, the authors report memory retention close to the individually-trained upper bound and higher PSNR/SSIM on the unseen SPA-data benchmark than the prior continual de-raining methods they compare against. They also introduce a similarity-based speedup that cuts training time by nearly half without hurting the results, and scalability mechanisms that keep GAN training and replay costs from growing linearly with the number of datasets.

What carries the argument

The load-bearing mechanism is GAN-based generative replay combined with consistency distillation. Each incoming dataset is memorized by a rain-generation GAN (modeled after [55], whose inference network predicts the parameters of a latent variable and whose generator renders the rain streak layer); replay images are produced by sampling the latent prior, generating a rain layer, and compositing it onto a clean background sampled from the current dataset, which makes the replayed data automatically paired with clean targets. The de-raining network is then optimized on interleaved new and replayed samples with the replay loss and the new loss weighted equally, and an L1 consistency loss on replayed images between the previous and current network enforces that current behavior on remembered inputs stays close to earlier behavior. A secondary mechanism, the similarity-based speedup, computes HOG descriptors and KL divergences between the new dataset and each replayed subset, maps the minimum divergence to [0,1], and scales the training iterations for the new dataset accordingly.

What would settle it

Replace one dataset's GAN in the stream with a deliberately poor one (e.g., trained on 10% of the images, or with its latent-to-rain mapping corrupted) and check whether memory retention for that dataset and generalization to SPA-data drop; alternatively, compute the distribution distance (such as FID) between GAN-generated rain layers and the dataset's real rain layers and verify that datasets with larger distances show more forgetting.

Watch

Extended reading notes

Core claim

The central claim is that generative replay enabled by per-dataset rain-generation GANs prevents catastrophic forgetting and improves cross-dataset generalization in image de-raining. Concretely, when a new de-raining dataset arrives, the framework trains a VRGNet-style GAN to capture that dataset's rain layer distribution, then forms a replayed dataset by sampling latent codes, generating rain layers, and adding them to clean backgrounds from the current dataset. The de-raining network is trained on the interleaving of new and replayed pairs with its own loss functions, while an L1 consistency loss between the outputs of the previous and current networks on the same replayed images aligns the new knowledge with the old. The paper reports that this procedure matches the memory performance of individually training on each dataset and outperforms PIGWM, NR, and DPL on unseen real-world SPA-data, with generalization continuing to improve as the dataset stream lengthens. It further claims that the same framework works across MFDNet, Restormer, and MPRNet, and that the similarity-based speedup preserves performance while reducing training iterations by an average of 44%.

Load-bearing premise

The framework assumes each per-dataset GAN faithfully captures the rain-streak distribution of its dataset, so that replay images formed by adding generated rain layers to clean backgrounds from the current dataset preserve enough prior knowledge for interleaved training; the paper does not directly measure replay fidelity.

Editorial extensions

If this is right

  • After a six-dataset stream, CLGID's average-memory PSNR stays within about 0.3 dB of the individually-trained reference, while sequential fine-tuning and the compared continual methods drop several dB.
  • Generalization on unseen SPA-data keeps rising as datasets are added (for MFDNet, 34.36 dB after six datasets), while competitors' generalization plateaus or declines.
  • The gains replicate across three de-raining backbones (MFDNet, Restormer, MPRNet), so the mechanism is not architecture-specific.
  • The similarity-based speedup cuts total training iterations by about 44% and total training time by 42% with negligible loss in memory or generalization.
  • Selective GAN training and replay-data reuse reduce GAN training/storage cost by over 50% and replay inference calls by 26.9%, with the reuse variant nearly matching full CLGID.

Reading between the lines

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

  • Because the replay mechanism only needs a generator of the corruption layer plus paired clean targets, the same continual-learning recipe should transfer to other restoration tasks such as deblurring, denoising, or raindrop removal; a testable prediction is that the consistency loss on replayed corruptions reduces forgetting there as well.
  • The consistency distillation effectively constrains the current network on the replay manifold, which implies that GAN replay fidelity (how closely generated rain layers resemble real ones) should predict memory retention; the paper does not measure this, so it is a direct check of the mechanism.
  • The generalization claim rests on a single unseen real-world test set (SPA-data); evaluating on additional real rain benchmarks with different acquisition conditions would test whether the reported gains are general or benchmark-specific.
  • The similarity-based speedup uses the minimum KL divergence between the new dataset and replayed subsets; if dataset order matters, reordering similar datasets adjacent should change the speedup magnitude while final performance stays stable, which is an easily testable consequence.
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

2 major / 6 minor

Summary. The paper presents CLGID, a continual learning framework for single-image de-raining that combines per-dataset GANs as generative memory, interleaved training on replayed and new data, and a consistency distillation loss. The framework is evaluated with three backbones (MFDNet, Restormer, MPRNet) on four-dataset and six-dataset streams, with memory evaluated on the training datasets and generalization on SPA-data and Real-Internet. The paper also proposes a HOG/KL-based similarity speedup, selective GAN training, and a replay-data-reuse mechanism. The main empirical claim is that CLGID retains past-dataset performance near the Individual upper bound and achieves the best SPA-data generalization among SF, PIGWM, NR, and DPL.

Significance. If the claims hold, CLGID would be a useful contribution to continual learning for image restoration, showing that generative replay plus distillation transfers across three backbones and several dataset orders. The paper deserves credit for reporting a 10-run stability analysis (Table IX), testing six datasets and multiple sequence orders, and for sharing a speedup variant CLGID† that reports roughly 44% fewer training iterations. However, the central generalization claim is weakened by the fact that the λ hyperparameter was selected on SPA-data, and the Appendix B proof of logarithmic replay cost is mathematically incorrect. The strengths are substantial, but the evaluation protocol and the proof need correction before the claims can be accepted.

major comments (2)
  1. [§IV-D, Table VIII] The SPA-data generalization comparison is not a clean holdout because the consistency-loss weight λ was selected by a 'comprehensive search' on the 1400-100L-1200M-100H sequence, and Table VIII reports the SPA-data column of that search. The same sequence's SPA-data results for MPRNet appear in Table III. Thus SPA-data has been used for model selection for CLGID, whereas the comparison methods' hyperparameters were not selected on SPA-data, biasing the comparison in CLGID's favor. Please re-select λ on a validation set that excludes SPA-data (or fix λ a priori) and re-run Tables I-V, or otherwise demonstrate that the ordering on SPA-data is robust to λ selection without using SPA-data. In particular, report the SPA-data performance of CLGID for the λ values tested in Table VIII and show that λ=1 is not chosen because it maximizes SPA-data on that sequence.
  2. [Appendix B, Eq. (21)-(22)] The proof that the replay-data-reuse total cost is C_N = O(M log N) is incorrect. The argument asserts that ε_n = max(0, M_max/(n-1) - M_min/(n-2)) vanishes for sufficiently large n; however, for M_max > M_min, this expression is positive for all large n and behaves like (M_max - M_min)/n. Consequently (n-2)ε_n does not tend to zero and the sum ∑_n (n-2)ε_n grows linearly in N, not to a constant. The claimed O(M log N) bound is therefore not established for datasets of unequal size. Please correct the proof, state explicitly if the bound requires all M_n equal, or revise the complexity claim. Additionally, the reuse mechanism replaces the Eq. (7) background images with backgrounds from earlier datasets; the paper should clarify whether this changes the replayed data distribution and why it does not hurt the distillation/replay objectives.
minor comments (6)
  1. [Tables I-III] The tables are titled 'Qualitative comparison' but report quantitative PSNR/SSIM values; rename them to 'Quantitative comparison'.
  2. [Tables I-V] The main tables report single-run results. Table IX gives stability for only one sequence; please report mean ± std for the main four- and six-dataset results, or at least for the key sequences, so the reader can judge whether the reported gaps are within noise.
  3. [§IV-B] No joint-training baseline is reported; adding a row that trains on the union of all datasets would help calibrate the cost of the continual-learning setting versus a non-continual upper bound.
  4. [§III-C] The threshold T_hat is set to 0.4 without a sensitivity study; please include an ablation over T_hat to justify the choice.
  5. [§III-B, Eq. (15)] The set {s_n_1, s_n_1, ..., s_n_{n-1}} has a duplicated index; it should read {s_n_1, s_n_2, ..., s_n_{n-1}}.
  6. [§III-A] The replay-fidelity assumption (that the per-dataset GANs faithfully represent rain-streak distributions) is not directly measured; a quantitative fidelity check (e.g., FID between generated and real rain layers, or a human study) would strengthen the hippocampal-replay analogy.

Circularity Check

1 steps flagged · score 5.0 of 10

SPA-data is used to select λ in Table VIII, so the 'unseen' SPA-data generalization claim is not a clean holdout; no derivation-level circularity otherwise.

  1. fitted input called prediction [Section IV-A (Experimental Settings) and Section IV-D (Validation on hyper-parameter λ), Table VIII]
    "For assessing generalization performance, following [20], we choose SPA-data [68] and Real-Internet [68] for qualitative and quantitative evaluation. These two datasets consist of real-world data that has never been seen during training. ... Considering the trade-off, we ultimately set λ to 1 through a comprehensive search, thereby achieving optimal memory and generalization performance."

    The paper's central generalization claim is measured on SPA-data, which is described as 'never seen during training.' However, Table VIII reports a λ sweep on the 1400-100L-1200M-100H sequence and includes SPA-data columns, and the text states that λ=1 was chosen 'through a comprehensive search, thereby achieving optimal memory and generalization performance.' That λ is then used in the main experiments whose SPA-data scores are presented as generalization to unseen data. SPA-data therefore participated in model selection for CLGID, while the compared methods' hyperparameters were not (per the paper) selected on SPA-data. Reporting these SPA-data results as an 'unseen' prediction is a fitted input presented as a prediction, biasing the headline generalization comparison upward.

full rationale

The core CLGID derivation is not circular: replay data are generated by independently trained per-dataset GANs (Eqs. 5-7), the de-raining network is trained on interleaved real and replayed data (Eqs. 8-10), and distillation (Eq. 11) is a standard consistency regularizer against the model's own previous state. None of these equations is defined in terms of the SPA-data outcome. The SPA-data test set is, however, used in model selection: Section IV-D reports a λ sweep on the 1400-100L-1200M-100H sequence, with SPA-data listed in Table VIII, and λ=1 is chosen 'through a comprehensive search' for 'optimal memory and generalization performance.' Since λ=1 is then used in the main results whose SPA-data columns are presented as generalization to data 'never seen during training,' the headline generalization comparison is not a clean holdout for CLGID relative to baselines. This is evaluation leakage and a fitted input called a prediction, rather than derivation-level circularity. The Appendix B proof of O(M log N) appears to rely on ε_n vanishing when dataset sizes differ, which is a correctness issue in a scalability claim, not a circularity. Overall score reflects the load-bearing but narrow contamination of the central generalization evidence.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim depends on the adequacy of GAN replay, the reliability of the previous network as distillation target, and the similarity heuristic for training speedup. Two hyperparameters are hand-set; no new physical entities are introduced.

free parameters (2)
  • lambda (consistency loss weight) = 1
    Chosen by a comprehensive search on the 1400-100L-1200M-100H sequence with MPRNet (Table VIII); that same sequence also appears among the main reported results, so part of the evaluation is mildly tuned.
  • T_hat (GAN training threshold) = 0.4
    Threshold for deciding when to train a new GAN (Eq. 18); set without an ablation or a principled derivation.
assumptions (5)
  • domain assumption Per-dataset GANs capture rain streak characteristics well enough for replay.
    Eqs. 1-7; no direct fidelity evaluation, central to knowledge retention.
  • domain assumption Generated rain layers added to the current dataset's clean backgrounds form valid training pairs.
    Eq. 7; cross-dataset composites could introduce a domain shift not present in real rain.
  • ad hoc to paper HOG feature histograms and KL divergence measure rain-pattern similarity relevant to training iterations.
    Eqs. 13-16; no evidence that lower KL corresponds to easier or faster learning.
  • domain assumption The previous de-raining network F_{n-1} is a reliable teacher for old knowledge.
    Eq. 11; if F_{n-1} already forgot or drifted, distillation reinforces the drift.
  • domain assumption Training on more synthetic rain datasets monotonically improves real-world generalization.
    Motivation in Section I and Figure 2; supported by experiments but not guaranteed in general.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Better De-raining Generalization via Rainy Characteristics Memorization and Replay." pith.science (2026). https://pith.science/paper/FWEKCIO4

@misc{pith2026250602477,
  author       = {Pith},
  title        = {Pith review of: Towards Better De-raining Generalization via Rainy Characteristics Memorization and Replay},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWEKCIO4}},
  note         = {Machine review of arXiv:2506.02477}
}
read the original abstract

Current image de-raining methods primarily learn from a limited dataset, leading to inadequate performance in varied real-world rainy conditions. To tackle this, we introduce a new framework that enables networks to progressively expand their de-raining knowledge base by tapping into a growing pool of datasets, significantly boosting their adaptability. Drawing inspiration from the human brain's ability to continuously absorb and generalize from ongoing experiences, our approach borrow the mechanism of the complementary learning system. Specifically, we first deploy Generative Adversarial Networks (GANs) to capture and retain the unique features of new data, mirroring the hippocampus's role in learning and memory. Then, the de-raining network is trained with both existing and GAN-synthesized data, mimicking the process of hippocampal replay and interleaved learning. Furthermore, we employ knowledge distillation with the replayed data to replicate the synergy between the neocortex's activity patterns triggered by hippocampal replays and the pre-existing neocortical knowledge. This comprehensive framework empowers the de-raining network to amass knowledge from various datasets, continually enhancing its performance on previously unseen rainy scenes. Our testing on three benchmark de-raining networks confirms the framework's effectiveness. It not only facilitates continuous knowledge accumulation across six datasets but also surpasses state-of-the-art methods in generalizing to new real-world scenarios.

Figures

Figures reproduced from arXiv: 2506.02477 by the authors.

Figure 1
Figure 1. Visual comparison between MFDNet [17], trained on a fixed de-raining dataset, and employing our CLGID to accumulate de-raining knowledge from increasingly abundant datasets. The red check mark indicates the network’s promis￾ing performance. While MFDNet trained on a fixed dataset can only handle specific types of real-world rain streaks (i.e., rain100H→Heavy, rain100L→Light), our CLGID is capable of addressing vario… view at source ↗
Figure 2
Figure 2. (a) Memory performance on individual datasets after [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Flowchart of the proposed CLGID framework. Given each incoming dataset [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Generalization performance variance on unseen SPA [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Generalization performance variance on unseen SPA [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Generalization performance variance on unseen SPA-data during training on a stream of six datasets. We showcase the [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Training iteration ratios of CLGID† compared with CLGID training on a stream of four datasets across four sequences. Each data point in the plots represents the similarity calculated between the current dataset and the previously learned dataset with the same color. (a…
Figure 9
Figure 9. Figure 9: Comparison of GAN training, storage, and replay costs between CLGID and its variants, CLGID [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Training loss curve under the sequence 1400-1200M-100H-100 using MPRNet. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Visual quality comparisons of different methods on SPA-data [ [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 65 canonical work pages

  1. [1]

    Generalized uav object detection via frequency do- main disentanglement,

    K. Wang, X. Fu, Y . Huang, C. Cao, G. Shi, and Z.-J. Zha, “Generalized uav object detection via frequency do- main disentanglement,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    work page 2023

  2. [2]

    Towards generalized uav object detection: A novel perspective from frequency domain disentanglement,

    K. Wang, X. Fu, C. Ge, C. Cao, and Z.-J. Zha, “Towards generalized uav object detection: A novel perspective from frequency domain disentanglement,”International Journal of Computer Vision, pp. 1–29, 2024

    work page 2024

  3. [3]

    Centernet-auto: A multi-object visual detection algo- rithm for autonomous driving scenes based on improved centernet,

    H. Wang, Y . Xu, Z. Wang, Y . Cai, L. Chen, and Y . Li, “Centernet-auto: A multi-object visual detection algo- rithm for autonomous driving scenes based on improved centernet,”IEEE Transactions on Emerging Topics in Computational Intelligence, 2023

    work page 2023

  4. [4]

    Rssformer: Foreground saliency enhance- ment for remote sensing land-cover segmentation,

    R. Xu, C. Wang, J. Zhang, S. Xu, W. Meng, and X. Zhang, “Rssformer: Foreground saliency enhance- ment for remote sensing land-cover segmentation,”IEEE Transactions on Image Processing, 2023

    work page 2023

  5. [5]

    Instance segmentation of biological images using graph convolutional network,

    R. Xu, Y . Li, C. Wang, S. Xu, W. Meng, and X. Zhang, “Instance segmentation of biological images using graph convolutional network,”Engineering Applications of Ar- tificial Intelligence, 2022

    work page 2022

  6. [6]

    Dc- net: Dual context network for 2d medical image seg- mentation,

    R. Xu, C. Wang, S. Xu, W. Meng, and X. Zhang, “Dc- net: Dual context network for 2d medical image seg- mentation,” inMedical Image Computing and Computer Assisted Intervention, 2021

    work page 2021

  7. [7]

    Uni-navid: A video- based vision-language-action model for unifying embod- ied navigation tasks,

    J. Zhang, K. Wang, S. Wang, M. Li, H. Liu, S. Wei, Z. Wang, Z. Zhang, and H. Wang, “Uni-navid: A video- based vision-language-action model for unifying embod- ied navigation tasks,”arXiv preprint arXiv:2412.06224, 2024

    arXiv 2024

  8. [8]

    Navid: Video-based vlm plans the next step for vision-and-language navigation,

    J. Zhang, K. Wang, R. Xu, G. Zhou, Y . Hong, X. Fang, Q. Wu, Z. Zhang, and H. Wang, “Navid: Video-based vlm plans the next step for vision-and-language navigation,” arXiv preprint arXiv:2402.15852, 2024

    arXiv 2024

Show all 68 references
  1. [9]

    Clearing the skies: A deep network architecture for single-image rain removal,

    X. Fu, J. Huang, X. Ding, Y . Liao, and J. Paisley, “Clearing the skies: A deep network architecture for single-image rain removal,”IEEE Transactions on Image Processing, 2017

    2017

  2. [10]

    Multi-stage progressive image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Multi-stage progressive image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    2021

  3. [11]

    Pro- gressive image deraining networks: A better and simpler baseline,

    D. Ren, W. Zuo, Q. Hu, P. Zhu, and D. Meng, “Pro- gressive image deraining networks: A better and simpler baseline,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 15

    2019

  4. [12]

    Multi-scale progressive fusion net- work for single image deraining,

    K. Jiang, Z. Wang, P. Yi, C. Chen, B. Huang, Y . Luo, J. Ma, and J. Jiang, “Multi-scale progressive fusion net- work for single image deraining,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020

    2020

  5. [13]

    Rethinking image deraining via rain streaks and vapors,

    Y . Wang, Y . Song, C. Ma, and B. Zeng, “Rethinking image deraining via rain streaks and vapors,” inEuropean Conference on Computer Vision, 2020

    2020

  6. [14]

    Heavy rain image restoration: Integrating physics model and conditional adversarial learning,

    R. Li, L.-F. Cheong, and R. T. Tan, “Heavy rain image restoration: Integrating physics model and conditional adversarial learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    2019

  7. [15]

    Multi-scale hybrid fusion net- work for single image deraining,

    K. Jiang, Z. Wang, P. Yi, C. Chen, G. Wang, Z. Han, J. Jiang, and Z. Xiong, “Multi-scale hybrid fusion net- work for single image deraining,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 7, pp. 3594–3608, 2021

    2021

  8. [16]

    Contrastive unfolding deraining network,

    Z. Huang and J. Zhang, “Contrastive unfolding deraining network,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 4, pp. 5155–5169, 2022

    2022

  9. [17]

    Multi-scale fusion and decomposition network for single image deraining,

    Q. Wang, K. Jiang, Z. Wang, W. Ren, J. Zhang, and C.- W. Lin, “Multi-scale fusion and decomposition network for single image deraining,”IEEE Transactions on Image Processing, 2023

    2023

  10. [18]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,”Neural Networks, 2019

    2019

  11. [19]

    Image de-raining via continual learning,

    M. Zhou, J. Xiao, Y . Chang, X. Fu, A. Liu, J. Pan, and Z.-J. Zha, “Image de-raining via continual learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    2021

  12. [20]

    Improv- ing de-raining generalization via neural reorganization,

    J. Xiao, M. Zhou, X. Fu, A. Liu, and Z.-J. Zha, “Improv- ing de-raining generalization via neural reorganization,” inProceedings of the IEEE/CVF International Confer- ence on Computer Vision, 2021

    2021

  13. [21]

    Dual prompt learning for continual rain removal from single images,

    M. Liu, W. Yang, Y . Hu, and J. Liu, “Dual prompt learning for continual rain removal from single images,” inProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 2023

    2023

  14. [22]

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory

    J. L. McClelland, B. L. McNaughton, and R. C. O’Reilly, “Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory.”Psychological Review, 1995

    1995

  15. [23]

    What learning systems do intelligent agents need? comple- mentary learning systems theory updated,

    D. Kumaran, D. Hassabis, and J. L. McClelland, “What learning systems do intelligent agents need? comple- mentary learning systems theory updated,”Trends in Cognitive Sciences, 2016

    2016

  16. [24]

    Hippocampal synaptic enhancement and information storage within a distributed memory system,

    B. L. McNaughton and R. G. Morris, “Hippocampal synaptic enhancement and information storage within a distributed memory system,”Trends in Neurosciences, 1987

    1987

  17. [25]

    Catastrophic forgetting in connectionist networks,

    R. M. French, “Catastrophic forgetting in connectionist networks,”Trends in Cognitive Sciences, 1999

    1999

  18. [26]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  19. [27]

    Incorporating rapid neocortical learn- ing of new schema-consistent information into comple- mentary learning systems theory

    J. L. McClelland, “Incorporating rapid neocortical learn- ing of new schema-consistent information into comple- mentary learning systems theory.”Journal of Experimen- tal Psychology: General, 2013

    2013

  20. [28]

    Schema- dependent gene activation and memory encoding in neo- cortex,

    D. Tse, T. Takeuchi, M. Kakeyama, Y . Kajii, H. Okuno, C. Tohyama, H. Bito, and R. G. Morris, “Schema- dependent gene activation and memory encoding in neo- cortex,”Science, 2011

    2011

  21. [29]

    Single- image deraining using an adaptive nonlocal means fil- ter,

    J.-H. Kim, C. Lee, J.-Y . Sim, and C.-S. Kim, “Single- image deraining using an adaptive nonlocal means fil- ter,” in2013 IEEE International Conference on Image Processing, 2013

    2013

  22. [30]

    A generalized low-rank appearance model for spatio-temporally correlated rain streaks,

    Y .-L. Chen and C.-T. Hsu, “A generalized low-rank appearance model for spatio-temporally correlated rain streaks,” inProceedings of the IEEE International Con- ference on Computer Vision, 2013

    2013

  23. [31]

    Transformed low-rank model for line pattern noise removal,

    Y . Chang, L. Yan, and S. Zhong, “Transformed low-rank model for line pattern noise removal,” inProceedings of the IEEE International Conference on Computer Vision, 2017

    2017

  24. [32]

    Automatic single- image-based rain streaks removal via image decomposi- tion,

    L.-W. Kang, C.-W. Lin, and Y .-H. Fu, “Automatic single- image-based rain streaks removal via image decomposi- tion,”IEEE Transactions on Image Processing, 2011

    2011

  25. [33]

    Self-learning based image decomposition with applica- tions to single image denoising,

    D.-A. Huang, L.-W. Kang, Y .-C. F. Wang, and C.-W. Lin, “Self-learning based image decomposition with applica- tions to single image denoising,”IEEE Transactions on Multimedia, 2013

    2013

  26. [34]

    Removing rain from a single image via discriminative sparse coding,

    Y . Luo, Y . Xu, and H. Ji, “Removing rain from a single image via discriminative sparse coding,” inProceedings of the IEEE International Conference on Computer Vi- sion, 2015

    2015

  27. [35]

    A hierarchical approach for rain or snow removing in a single color image,

    Y . Wang, S. Liu, C. Chen, and B. Zeng, “A hierarchical approach for rain or snow removing in a single color image,”IEEE Transactions on Image Processing, 2017

    2017

  28. [36]

    Syn2real transfer learning for image deraining using gaussian processes,

    R. Yasarla, V . A. Sindagi, and V . M. Patel, “Syn2real transfer learning for image deraining using gaussian processes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020

    2020

  29. [37]

    Semi- supervised transfer learning for image rain removal,

    W. Wei, D. Meng, Q. Zhao, Z. Xu, and Y . Wu, “Semi- supervised transfer learning for image rain removal,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    2019

  30. [38]

    Unpaired learning for deep image deraining with rain direction regularizer,

    Y . Liu, Z. Yue, J. Pan, and Z. Su, “Unpaired learning for deep image deraining with rain direction regularizer,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021

    2021

  31. [39]

    Unsupervised deraining: Where contrastive learning meets self-similarity,

    Y . Ye, C. Yu, Y . Chang, L. Zhu, X.-l. Zhao, L. Yan, and Y . Tian, “Unsupervised deraining: Where contrastive learning meets self-similarity,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    2022

  32. [40]

    Un- supervised single image deraining with self-supervised constraints,

    X. Jin, Z. Chen, J. Lin, Z. Chen, and W. Zhou, “Un- supervised single image deraining with self-supervised constraints,” inIEEE International Conference on Image Processing, 2019

    2019

  33. [41]

    Single-image IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 16 deraining via recurrent residual multiscale networks,

    Y . Zheng, X. Yu, M. Liu, and S. Zhang, “Single-image IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 16 deraining via recurrent residual multiscale networks,” IEEE transactions on neural networks and learning systems, vol. 33, no. 3, pp. 1310–1323, 2020

    2020

  34. [42]

    Direction and residual awareness curriculum learning network for rain streaks removal,

    Y . Chang, M. Chen, C. Yu, Y . Li, L. Chen, and L. Yan, “Direction and residual awareness curriculum learning network for rain streaks removal,”IEEE transactions on neural networks and learning systems, 2023

    2023

  35. [43]

    Rain streaks removal for single image via kernel-guided convolutional neural network,

    Y .-T. Wang, X.-L. Zhao, T.-X. Jiang, L.-J. Deng, Y . Chang, and T.-Z. Huang, “Rain streaks removal for single image via kernel-guided convolutional neural network,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 8, pp. 3664–3676, 2020

    2020

  36. [44]

    All- in-one image restoration for unknown corruption,

    B. Li, X. Liu, P. Hu, Z. Wu, J. Lv, and X. Peng, “All- in-one image restoration for unknown corruption,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17 452–17 462

    2022

  37. [45]

    Lightweight pyramid networks for image deraining,

    X. Fu, B. Liang, Y . Huang, X. Ding, and J. Paisley, “Lightweight pyramid networks for image deraining,” IEEE transactions on neural networks and learning systems, vol. 31, no. 6, pp. 1794–1807, 2019

    2019

  38. [46]

    Clearer: Multi-scale neural architecture search for image restora- tion,

    Y . Gou, B. Li, Z. Liu, S. Yang, and X. Peng, “Clearer: Multi-scale neural architecture search for image restora- tion,”Advances in neural information processing sys- tems, vol. 33, pp. 17 129–17 140, 2020

    2020

  39. [47]

    Mair: A locality-and continuity-preserving mamba for image restoration,

    B. Li, H. Zhao, W. Wang, P. Hu, Y . Gou, and X. Peng, “Mair: A locality-and continuity-preserving mamba for image restoration,”arXiv preprint arXiv:2412.20066, 2024

    2024 arXiv

  40. [48]

    Scale-free single image deraining via visibility-enhanced recurrent wavelet learning,

    W. Yang, J. Liu, S. Yang, and Z. Guo, “Scale-free single image deraining via visibility-enhanced recurrent wavelet learning,”IEEE Transactions on Image Processing, 2019

    2019

  41. [49]

    Dual-path coupled image deraining network via spatial- frequency interaction,

    Y . He, A. Jiang, L. Jiang, L. Peng, Z. Wang, and L. Wang, “Dual-path coupled image deraining network via spatial- frequency interaction,” in2024 IEEE International Con- ference on Image Processing (ICIP). IEEE, 2024, pp. 1452–1458

    2024

  42. [50]

    Single image deraining using bilateral recurrent network,

    D. Ren, W. Shang, P. Zhu, Q. Hu, D. Meng, and W. Zuo, “Single image deraining using bilateral recurrent network,”IEEE Transactions on Image Processing, 2020

    2020

  43. [51]

    Removing arbitrary- scale rain streaks via fractal band learning with self- supervision,

    W. Yang, S. Wang, and J. Liu, “Removing arbitrary- scale rain streaks via fractal band learning with self- supervision,”IEEE Transactions on Image Processing, 2020

    2020

  44. [52]

    An efficient single image de-raining model with decoupled deep networks,

    W. Li, G. Chen, and Y . Chang, “An efficient single image de-raining model with decoupled deep networks,”IEEE Transactions on Image Processing, vol. 33, pp. 69–81, 2023

    2023

  45. [53]

    Dual attention- in-attention model for joint rain streak and raindrop removal,

    K. Zhang, D. Li, W. Luo, and W. Ren, “Dual attention- in-attention model for joint rain streak and raindrop removal,”IEEE Transactions on Image Processing, 2021

    2021

  46. [54]

    A dual cnn architec- ture for single image raindrop and rain streak removal,

    A. Sivaanpu and K. Thanikasalam, “A dual cnn architec- ture for single image raindrop and rain streak removal,” in2022 7th International Conference on Information Technology Research (ICITR). IEEE, 2022, pp. 1–6

    2022

  47. [55]

    From rain generation to rain removal,

    H. Wang, Z. Yue, Q. Xie, Q. Zhao, Y . Zheng, and D. Meng, “From rain generation to rain removal,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    2021

  48. [56]

    Learning gated non-local residual for single- image rain streak removal,

    L. Zhu, Z. Deng, X. Hu, H. Xie, X. Xu, J. Qin, and P.- A. Heng, “Learning gated non-local residual for single- image rain streak removal,”IEEE Transactions on Cir- cuits and Systems for Video Technology, 2020

    2020

  49. [57]

    Pixel- learnable 3dlut with saturation-aware compensation for image enhancement,

    J. Liu, Q. Li, X. Min, Y . Su, G. Zhai, and X. Yang, “Pixel- learnable 3dlut with saturation-aware compensation for image enhancement,”IEEE Transactions on Multimedia, 2024

    2024

  50. [58]

    Single-image real-time rain removal based on depth- guided non-local features,

    X. Hu, L. Zhu, T. Wang, C.-W. Fu, and P.-A. Heng, “Single-image real-time rain removal based on depth- guided non-local features,”IEEE Transactions on Image Processing, 2021

    2021

  51. [59]

    Rcdnet: An interpretable rain convolutional dictionary network for single image deraining,

    H. Wang, Q. Xie, Q. Zhao, Y . Li, Y . Liang, Y . Zheng, and D. Meng, “Rcdnet: An interpretable rain convolutional dictionary network for single image deraining,”IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  52. [60]

    A comprehensive benchmark analysis of single image deraining: Current challenges and future perspectives,

    S. Li, W. Ren, F. Wang, I. B. Araujo, E. K. Tokuda, R. H. Junior, R. M. Cesar-Jr, Z. Wang, and X. Cao, “A comprehensive benchmark analysis of single image deraining: Current challenges and future perspectives,” International Journal of Computer Vision, 2021

    2021

  53. [61]

    Single image deraining: From model-based to data-driven and beyond,

    W. Yang, R. T. Tan, S. Wang, Y . Fang, and J. Liu, “Single image deraining: From model-based to data-driven and beyond,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020

    2020

  54. [62]

    Incremental image de-raining via associative memory,

    Y . Gu, C. Wang, and J. Li, “Incremental image de-raining via associative memory,” inProceedings of the AAAI Conference on Artificial Intelligence, 2023

    2023

  55. [63]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2005

    2005

  56. [64]

    On information and sufficiency,

    S. Kullback and R. A. Leibler, “On information and sufficiency,”The Annals of Mathematical Statistics, 1951

    1951

  57. [65]

    Deep joint rain detection and removal from a single image,

    W. Yang, R. T. Tanet al., “Deep joint rain detection and removal from a single image,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017

    2017

  58. [66]

    Removing rain from single images via a deep detail network,

    X. Fu, J. Huanget al., “Removing rain from single images via a deep detail network,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017

    2017

  59. [67]

    Density-aware single im- age de-raining using a multi-stream dense network,

    H. Zhang and V . M. Patel, “Density-aware single im- age de-raining using a multi-stream dense network,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018

    2018

  60. [68]

    Spatial attentive single-image deraining with a high quality real rain dataset,

    T. Wang, X. Yang, K. Xu, S. Chen, Q. Zhang, and R. W. Lau, “Spatial attentive single-image deraining with a high quality real rain dataset,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    2019

Pith tools

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