Pith. sign in

REVIEW 4 major objections 6 minor 55 references

CLIP-driven rain perception: Adaptive deraining with pattern-aware network routing and mask-guided cross-attention

T0 review · 4 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 A CLIP-based router sends each rainy image to a specialist deraining sub-network, and the paper reports that this beat every compared method on mixed rain benchmarks.

desk verdict The CLIP-routing idea is fresh, but the SOTA numbers are tuned on the test set and the routing is never shown to actually perceive rain types; a solid submission for review, not a pass as-is. read the letter →

arxiv 2506.01366 v1 pith:3ZT2XQZN submitted 2025-06-02 cs.CV

Cong Guan , Osamu Yoshie This is my paper
classification cs.CV
keywords imagederainingCLIPrainpatternperceptionadaptiveroutingmask-guidedcross-attentiondynamiclossschedulingmixeddatasettransformerrestoration
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

Deraining is usually done by one network that tries to remove every kind of rain. The paper's claim is that this is the wrong design: rain differs enough in drop density, streak orientation, and intensity that the network should first recognize what pattern it is looking at and route the image to a branch specialized for that pattern. It uses CLIP's pre-trained image-text matching to do the recognition, scoring the image against prompts such as "very sparse raindrops" versus "dense raindrops" and choosing the branch with the highest score. Two further components handle the restoration itself: a mask-guided cross-attention module separates rainy and clean regions and lets them exchange information, and a dynamic loss schedule shifts optimization from low-frequency smooth areas early in training to high-frequency detail late in training. On Rain100L, Rain100H, Rain800, and their mixture, the complete system reports the highest PSNR/SSIM among the compared methods, with the clearest edge on the mixed dataset (32.42 dB PSNR versus 32.28 for the runner-up).

What carries the argument

The mechanism that carries the argument is the routing identity $s=\arg\max_i s_i$ with $s_i=e^{I_1\cdot T_i}/\sum_j e^{I_1\cdot T_j}$, computed from CLIP image and text embeddings; this single index decides which sub-network processes the image and is what turns the model into a divide-and-conquer system for rain types. The second mechanism is mask-guided cross-attention: a convolutional network predicts a confidence map $\hat{M}$, features are split as $F_r=F\odot\hat{M}$ and $F_n=F\odot(1-\hat{M})$, and cross-attention blocks exchange channel importance from the non-rainy side and spatial importance from the rainy side before the two streams are summed. The third is the dynamic loss $l_\tau(\varepsilon)=\varepsilon^{\beta+\eta f(\tau)/T}$, whose exponent rises across training to first favor low-frequency smooth regions and later high-frequency detail; the paper argues that $\ell^1$'s constant gradient and $\ell^2$'s error-proportional gradient both misalign with this natural training progression.

What would settle it

Run CLIP-RPN with the routing index replaced by a random permutation of the same branch assignments on a held-out mixed test set; if PSNR stays at 32.42, routing is not the cause. Separately, annotate held-out images with ground-truth rain density and measure how often CLIP's argmax prompt matches; near-chance accuracy would contradict the perception claim.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that rain-pattern perception can be reduced to a visual-language matching problem: a pre-trained CLIP model encodes the rainy image and a few text prompts, a softmax similarity score picks the best-matching prompt, and the argmax index selects one of several specialist sub-networks, so each branch is trained on a narrower slice of rain conditions rather than all of them at once. Inside the chosen branch, a small convolutional network predicts a rain-confidence mask, the features are split into rainy and non-rainy parts, and cross-attention lets the clear regions' channel importance guide the rainy regions while the rainy regions' spatial importance guides the clear regions. The training loss is also made time-dependent: its exponent grows during training so early gradients favor small errors in smooth regions and later gradients favor large errors in detail. The paper's ablations tie the mixed-dataset result to all three pieces: removing routing lowers PSNR from 32.42 to 32.31, removing mask-guided attention lowers it to 32.11, and replacing the scheduled loss with $\ell^1$, $\ell^2$, or Huber lowers it to 32.25, 32.18, or 32.19.

Load-bearing premise

The routing advantage rests on CLIP's zero-shot text-image scores correctly telling which rain pattern an image contains; if those scores are unreliable, images go to the wrong specialist branch and the divide-and-conquer gain disappears.

Editorial extensions

If this is right

  • If routing is what the experiments suggest, deraining systems can be assembled from a small set of specialist branches plus a router, rather than one monolithic network expected to absorb all rain variation.
  • The mask-guided interaction implies that explicitly separating affected and unaffected regions, and letting one guide the other, is a transferable inductive bias for restoration, not just an auxiliary supervision trick.
  • The dynamic loss schedule implies that static l1/l2/Huber losses leave performance on the table during long training runs; time-varying exponents could be dropped into other restoration pipelines.
  • Because the gains concentrate on Rain800 and the mixed dataset, the paper's protocol says benchmark mixtures, not single-pattern datasets, are where deraining methods should be compared.

Reading between the lines

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

  • The paper does not directly measure whether CLIP's routing decisions match human labels of rain type; a natural test would be to score routing accuracy on held-out images with annotated density and orientation and see whether correct routing, not just final PSNR, explains the gain.
  • The component ablation suggests mask-guided attention carries at least as much of the improvement as the router; transplanting the attention module into a fixed single-branch derainer would show whether most of the gain survives without CLIP.
  • The same prompt-scoring trick could apply to other weather degradations such as haze or snow by swapping prompt sets, and the paper's conclusion names dehazing and snow removal as future work; testing router accuracy before fine-tuning each branch would make that transfer diagnostic.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes CLIP-RPN, a deraining network that uses CLIP text-image similarity scores to classify each input image into one of several rain-pattern classes and route it to a specialized sub-network. The method also introduces Mask-Guided Cross-Attention (MGCA), which uses a multi-level predicted rain mask to separate features into rainy and non-rainy regions and enables cross-attention between them, and a Dynamic Loss Scheduling (DLS) strategy whose loss exponent increases during training. Experiments on Rain100L, Rain100H, Rain800, and a mixed dataset report PSNR/SSIM improvements over several all-in-one restoration baselines, with the largest gains on Rain800 and the mixed set. Ablations attribute the gains to RPN and MGCA, and loss comparisons attribute gains to DLS.

Significance. If the reported gains are genuine and attributable to the proposed mechanisms, the paper would contribute a useful new way to apply CLIP to low-level vision tasks and a simple, transferable loss-scheduling idea. The strengths of the manuscript include the clear presentation of the divide-and-conquer routing concept, the multi-scale mask prediction and feature-interaction design, the explicit reporting of parameter counts and inference times, and a fairly broad set of ablations, visualizations, and loss comparisons. However, a load-bearing weakness is that the prompt set and the DLS hyperparameters were selected on the same test sets that produce the headline numbers, and the CLIP routing is never validated at the per-image level. Unless those issues are addressed, the central claim that semantic routing, rather than added parameters or arbitrary data partitioning, drives the performance gains is not established.

major comments (4)
  1. [§4.2, §4.4.1, Table 6, Figs. 7–8] The final results in Table 3 are obtained with Prompt 3, which is selected because it has the best PSNR/SSIM in Table 6 on exactly the same Rain100L/H, Rain800, and Mixed test sets, and with β=0.8, η=2.3 selected from the curves in Figs. 7–8, which appear to be evaluated on the same Mixed test set. This makes part of the reported state-of-the-art margin a test-set-fitted quantity. The authors should use a separate validation split for prompt and hyperparameter selection and report results on a true test split, or they should report repeated-run statistics with significance tests. Without this, the small differences in Table 3 (for example, 0.01 dB on Rain100L and 0.04 dB on Rain100H) cannot be distinguished from selection noise.
  2. [§3.2, Eq. (2), Tables 2 and 7–9] The routing mechanism's semantic correctness is not established. The dataset-level CLIP score distributions are mutually inconsistent and do not match the known properties of the datasets. For Rain100L, which the paper itself describes as light rain, Table 2 reports 46.00% light drizzle and 41.50% heavy downpour, while Table 7 assigns 67.50% of Rain100L images to 'intense rainfall', Table 8 assigns 67.50% to 'dense raindrops', and Table 9 assigns 70.50% to 'unnatural...poor quality'. These numbers change dramatically with prompt wording, which suggests that the argmax in Eq. (2) is influenced by prompt phrasing rather than by the physical rain type. The paper reports no per-image routing accuracy, confusion matrix, or comparison against random routing, so there is no evidence that images are sent to the branch corresponding to their actual rain pattern. The paper's own explanation that Prompt 1 performs worse because one branch receives only 0.22% of images on the Mixed set further indicates that routing distribution, rather than semantic correctness, may drive the differences.
  3. [§4.4.2, Table 10] The ablation does not isolate the routing mechanism from added capacity. Removing RPN changes more than routing: it also removes the extra branch parameters and the mask supervision; removing MGCA removes an entire attention module. The full-model vs no-RPN difference on Mixed is only 0.11 dB PSNR, and the full vs no-MGCA difference is 0.31 dB, yet no error bars or multiple runs are reported. A control with random assignment of images to the same number of sub-networks and an equal parameter budget should be compared; otherwise the 0.11 dB gain could be due to capacity or arbitrary partitioning. The claim at the end of §4.4.2 that 'rain pattern awareness and adaptive routing contribute significantly' is stronger than the evidence supports.
  4. [§4.2, Fig. 1(a), Table 5] The number of sub-networks instantiated in the experiments is not stated. Prompt 3 contains only two text prompts, while Fig. 1(a) shows three sub-networks and the method section refers to an unspecified n in Eq. (2). Since both the parameter count in Table 4 and the routing behavior depend on n, the paper must specify how many sub-networks are used for each prompt set. It should also state explicitly whether the frozen CLIP encoder parameters are included in the reported 32.72 M parameter count; CLIP ViT-B/32 contains far more parameters than 32.72 M, and the current text does not clarify whether Table 4 reports trainable parameters only.
minor comments (6)
  1. [Abstract] The sentence 'We also introduces a dynamic loss scheduling mechanism' should read 'We also introduce a dynamic loss scheduling mechanism'.
  2. [§4.4.2] The text says the ablation results are 'summarized in Table 11', but the corresponding table is numbered Table 10; this cross-reference should be fixed.
  3. [Table 5] Prompt 3 contains a typo: 'distoration' should be 'distortion'.
  4. [Table 3] It is not stated whether the baseline numbers in Table 3 are copied from the original papers or re-produced under the same training protocol; this information is needed for a fair comparison.
  5. [§4.4.5, Fig. 9] The visualization of predicted masks is helpful, but the caption does not explain the color scale or the difference between M1, M2, and M3 beyond the qualitative description in the text; a short legend would improve interpretability.
  6. [§5 and Data Availability] The conclusion claims the method is suitable for 'real-world applications', but all evaluations are on synthetic benchmarks; the claim should be softened or supplemented with real-world rainy image tests.

Circularity Check

2 steps flagged · score 5.0 of 10

The architectural derivation is not definitionally circular, but the reported best prompt and the dynamic-loss hyperparameters are selected on the test set, so parts of the final SOTA numbers reduce to fitted choices.

  1. fitted input called prediction [Section 4.4.1 (Table 6) and Section 4.2]
    "The quantitative results of our proposed method with these different prompts are presented in Table 6. It is evident that on most cases, Prompt 3 consistently outperforms the other two prompts across all datasets in terms of both PSNR and SSIM metrics. ... The default prompt used in the experiments is Prompt 3 from Table 5."

    Prompt 3 is selected because it achieves the highest PSNR/SSIM on the same public test sets that are later reported as the final results in Table 3. The choice is therefore a discrete parameter fitted to the evaluation benchmarks, and the statement 'Prompt 3 outperforms the other prompts' is the selection rule itself rather than an independent prediction. Any method that searches over prompt wording on the test set will, by construction, report the selected prompt's test score as its best result, so the final comparison against other methods inherits this test-set selection.

  2. fitted input called prediction [Section 4.4.4 (Figures 7-8) and Section 4.2]
    "Overall, the model achieves optimal performance when β and η are set to 0.8 and 2.3, respectively. ... In the comparison experiments, the β and η in the loss function were set to 0.8 and 2.3, respectively."

    Figures 7 and 8 sweep β and η while plotting the Mixed-dataset test PSNR, and the chosen values are the maximum points of those test curves. The same test-set numbers are then used in Tables 3 and 11 to demonstrate that the dynamic loss scheduling DLS outperforms l1, l2, and Huber losses. Consequently the reported superiority of DLS is partly forced by optimizing its two scalar hyperparameters on the evaluation set; the improvement is a fitted maximum rather than an out-of-sample prediction of the loss design.

full rationale

The core architecture is not defined in terms of its results: CLIP-RPN's routing mechanism (Eq. 2), mask-guided cross-attention (Eqs. 7-13), and dynamic loss scheduling (Eq. 14) are all explicitly constructed and then ablated in Table 10, so the main method sections do not reduce to the output. No load-bearing self-citation or imported uniqueness theorem appears, and the paper does not rename an existing result. The circularity is concentrated in two test-set selection steps: the prompt wording is chosen from Table 6 on the same benchmarks used for the final comparison, and the loss exponents β and η are chosen from Figures 7-8 on the same Mixed test set used in Tables 3 and 11. These choices make part of the reported SOTA margin a fitted quantity. The paper's evidence that CLIP routing is semantically meaningful is also weak and internally inconsistent (e.g., Table 9 assigns 70.5% of Rain100L, the light-rain dataset, to the 'unnatural and poor quality' prompt), but that is a correctness and validation concern rather than a definitional circularity. Overall, the architectural claims retain independent content, so the score is moderate rather than extreme.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on hand-chosen hyperparameters (mask threshold, BCE weight, beta, eta, and prompt wording) and on the adequacy of CLIP as a rain classifier. The prompt and loss-exponent choices were tuned on test data, which materially weakens the reported improvements. There are no fully invented physical entities; the 'rain pattern categories' are pragmatic classes derived from text prompts.

free parameters (5)
  • Mask threshold for ground-truth rain mask = 0.1
    Eq. (4) labels any pixel with |rainy - clean| > 0.1 as a rain pixel; the threshold is chosen by hand with no sensitivity analysis.
  • BCE loss weight = 0.1
    Eq. (15) multiplies each mask BCE loss by 0.1; stated as an 'empirical weight factor' with no derivation.
  • DLS exponent start parameter beta = 0.8
    Hyperparameter in Eq. (14) controlling the initial loss exponent; tuned on the Mixed test set (Fig. 7).
  • DLS exponent growth parameter eta = 2.3
    Hyperparameter in Eq. (14); tuned on the Mixed test set (Fig. 8).
  • Prompt set and prompt wording = Prompt 3 in Table 5
    Three GPT-4 generated prompt sets were compared, and the best on the test sets was chosen (Table 6), making the final prompt a fitted choice.
assumptions (3)
  • domain assumption CLIP similarity scores accurately distinguish rain patterns
    Used for routing in Eq. (2); only validated with the same prompt style later used in the final model, not an independent benchmark.
  • ad hoc to paper Pixel-wise mask threshold at 0.1 captures rain regions
    Defined in Eq. (4); no principled justification or sensitivity analysis.
  • domain assumption Networks recover low-frequency content before high-frequency details
    Motivates the dynamic loss in Section 3.4; the cited works [45,46] do not directly establish this specific scheduling benefit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLIP-driven rain perception: Adaptive deraining with pattern-aware network routing and mask-guided cross-attention." pith.science (2026). https://pith.science/paper/3ZT2XQZN

@misc{pith2026250601366,
  author       = {Pith},
  title        = {Pith review of: CLIP-driven rain perception: Adaptive deraining with pattern-aware network routing and mask-guided cross-attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZT2XQZN}},
  note         = {Machine review of arXiv:2506.01366}
}
abstract

Existing deraining models process all rainy images within a single network. However, different rain patterns have significant variations, which makes it challenging for a single network to handle diverse types of raindrops and streaks. To address this limitation, we propose a novel CLIP-driven rain perception network (CLIP-RPN) that leverages CLIP to automatically perceive rain patterns by computing visual-language matching scores and adaptively routing to sub-networks to handle different rain patterns, such as varying raindrop densities, streak orientations, and rainfall intensity. CLIP-RPN establishes semantic-aware rain pattern recognition through CLIP's cross-modal visual-language alignment capabilities, enabling automatic identification of precipitation characteristics across different rain scenarios. This rain pattern awareness drives an adaptive subnetwork routing mechanism where specialized processing branches are dynamically activated based on the detected rain type, significantly enhancing the model's capacity to handle diverse rainfall conditions. Furthermore, within sub-networks of CLIP-RPN, we introduce a mask-guided cross-attention mechanism (MGCA) that predicts precise rain masks at multi-scale to facilitate contextual interactions between rainy regions and clean background areas by cross-attention. We also introduces a dynamic loss scheduling mechanism (DLS) to adaptively adjust the gradients for the optimization process of CLIP-RPN. Compared with the commonly used $l_1$ or $l_2$ loss, DLS is more compatible with the inherent dynamics of the network training process, thus achieving enhanced outcomes. Our method achieves state-of-the-art performance across multiple datasets, particularly excelling in complex mixed datasets.

Figures

Figures reproduced from arXiv: 2506.01366 by the authors.

Figure 1
Figure 1. Overview of the proposed method 3.1. Overview The overall structure of the proposed network (CLIP-RPN), along with its detailed components, is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 3
Figure 3. The commonly used loss functions and their gradients with respect to the prediction error. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 2
Figure 2. The logarithmic gradients distribution (the second row) of natural images (the first row). [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The proposed loss function and its gradient with respect to the prediction error and the dynamically scheduled exponential parameter. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of different deraining models. For each group of images, the first row shows the original images, the second row shows the mean absolute error of the images. 16 [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Convergence curves of different losses. DLS achieves the highest PSNR and SSIM values across all datasets, particularly excelling in the challenging Mixed dataset. The convergence curves in [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Analysis of the impact of different β values on the performance of the proposed method. η is fixed at 2.3. 1.0 1.5 2.0 2.5 3.0 3.5 Value of ´ 32.27 32.30 32.33 32.35 32.38 32.40 32.42 PSNR PSNR (a) PSNR 1.0 1.5 2.0 2.5 3.0 3.5 Value of ´ 0.8665 0.8667 0.8670 0.8672 0.8…
Figure 8
Figure 8. Figure 8: Analysis of the impact of different η values on the performance of the proposed method. β is fixed at 0.8. 21 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Visualization of the predicted mask. To clearly demonstrate the effectiveness of mask-guided cross-attention, we visualize the predicted masks at each level of CLIP-RPN, as shown in [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 45 canonical work pages

  1. [1]

    X. Chen, J. Pan, J. Dong, J. Tang, Towards unified deep image deraining: A survey and a new benchmark, IEEE Trans. Pattern Anal. Mach. Intell. (2025) 1–20doi:10.1109/TPAMI.2025.3556133

    arXiv 2025

  2. [2]

    W. Yang, R. T. Tan, S. Wang, Y . Fang, J. Liu, Single image deraining: From model-based to data-driven and beyond, IEEE Trans. Pattern Anal. Mach. Intell. 43 (11) (2020) 4059–4077

    work page 2020

  3. [3]

    Zhang, Y

    Z. Zhang, Y . Wei, H. Zhang, Y . Yang, S. Yan, M. Wang, Data-driven single image deraining: A comprehensive review and new perspectives, Pattern Recog. 143 (2023) 109740. doi:https://doi.org/10.1016/j.patcog.2023.109740

    work page arXiv 2023

  4. [4]

    Z. Su, Y . Zhang, J. Shi, X.-P. Zhang, A survey of single image rain removal based on deep learning, ACM Comput. Surv. 56 (4) (2023) 1–35

    work page 2023

  5. [5]

    H. Wang, Y . Wu, M. Li, et al., Survey on rain removal from videos or a single image, Sci. China Inf. Sci. 65 (11) (2022) 111101. doi: 10.1007/s11432-020-3225-9

    work page doi:10.1007/s11432-020-3225-9 2022

  6. [6]

    S. Gu, D. Meng, W. Zuo, L. Zhang, Joint convolutional analysis and synthesis sparse representation for single image layer separation, in: Int. Conf. Comput. Vis., 2017

    work page 2017

  7. [7]

    Y . Li, R. T. Tan, X. Guo, J. Lu, M. S. Brown, Rain streak removal using layer priors, in: IEEE Conf. Comput. Vis. Pattern Recog., 2016. 23

    work page 2016

  8. [8]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet classification with deep convolutional neural networks, in: Adv. Neural Inform. Process. Syst., NIPS’12, Curran Associates Inc., Red Hook, NY , USA, 2012, p. 1097–1105

    work page 2012

Show all 55 references
  1. [9]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y . Bengio, Generative adversarial nets, in: Adv. Neural Inform. Process. Syst., NIPS’14, MIT Press, Cambridge, MA, USA, 2014, p. 2672–2680

    2014

  2. [10]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Adv. Neural Inform. Process. Syst. 30 (2017)

    2017

  3. [11]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  4. [12]

    J. Ho, A. Jain, P. Abbeel, Denoising di ffusion probabilistic models, Adv. Neural Inform. Process. Syst. 33 (2020) 6840–6851

    2020

  5. [13]

    W. Yang, R. T. Tan, J. Feng, J. Liu, Z. Guo, S. Yan, Deep joint rain detection and removal from a single image, in: IEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 1685–1694. doi:10.1109/CVPR.2017.183

    2017 doi

  6. [14]

    Zhang, V

    H. Zhang, V . Sindagi, V . M. Patel, Image de-raining using a conditional generative adversarial network, IEEE Trans. Circuits Syst. Video Technol. 30 (11) (2020) 3943–3956. doi:10.1109/TCSVT.2019.2920407

    2020

  7. [15]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: Int. Conf. Mach. Learn., PmLR, 2021, pp. 8748–8763

    2021

  8. [16]

    P. Li, J. Jin, G. Jin, L. Fan, X. Gao, T. Song, X. Chen, Deep scale-space mining network for single image deraining, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 4276–4285

    2022

  9. [17]

    J. Xu, W. Zhao, P. Liu, X. Tang, Removing rain and snow in a single image using guided filter, in: IEEE Int. Conf. Comput. Sci. Autom. Eng., V ol. 2, 2012, pp. 304–307.doi:10.1109/CSAE.2012.6272780

    2012

  10. [18]

    Fu, L.-W

    Y .-H. Fu, L.-W. Kang, C.-W. Lin, C.-T. Hsu, Single-frame-based rain removal via image decomposition, in: IEEE Int. Conf. Acoustics, Speech and Signal Process., 2011, pp. 1453–1456. doi:10.1109/ICASSP.2011.5946766

    2011

  11. [19]

    Zoran, Y

    D. Zoran, Y . Weiss, From learning models of natural image patches to whole image restoration, in: Int. Conf. Comput. Vis., 2011, pp. 479–486. doi:10.1109/ICCV.2011.6126278

    2011

  12. [20]

    S. Du, Y . Liu, M. Ye, Z. Xu, J. Li, J. Liu, Single image deraining via decorrelating the rain streaks and background scene in gradient domain, Pattern Recog. 79 (2018) 303–317

    2018

  13. [21]

    X. Fu, J. Huang, X. Ding, Y . Liao, J. Paisley, Clearing the skies: A deep network architecture for single-image rain removal, IEEE Trans. Image Process. 26 (6) (2017) 2944–2956

    2017

  14. [22]

    X. Fu, J. Huang, D. Zeng, Y . Huang, X. Ding, J. Paisley, Removing rain from single images via a deep detail network, in: IEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 3855–3863

    2017

  15. [23]

    Hu, C.-W

    X. Hu, C.-W. Fu, L. Zhu, P.-A. Heng, Depth-attentional features for single-image rain removal, in: IEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 8022–8031

    2019

  16. [24]

    X. Fu, J. Xiao, Y . Zhu, A. Liu, F. Wu, Z.-J. Zha, Continual image deraining with hypergraph convolutional networks, IEEE Trans. Pattern Anal. Mach. Intell. 45 (8) (2023) 9534–9551

    2023

  17. [25]

    Yasarla, V

    R. Yasarla, V . A. Sindagi, V . M. Patel, Syn2real transfer learning for image deraining using gaussian processes, in: IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 2726–2736

    2020

  18. [26]

    Y . Ye, C. Yu, Y . Chang, L. Zhu, X.-L. Zhao, L. Yan, Y . Tian, Unsupervised deraining: Where contrastive learning meets self-similarity, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 5821–5830

    2022

  19. [27]

    X. Chen, J. Pan, K. Jiang, Y . Li, Y . Huang, C. Kong, L. Dai, Z. Fan, Unpaired deep image deraining using dual contrastive learning, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 2017–2026

    2022

  20. [28]

    X. Chen, H. Li, M. Li, J. Pan, Learning a sparse transformer network for e ffective image deraining, in: IEEE Conf. Comput. Vis. Pattern Recog., 2023, pp. 5896–5905

    2023

  21. [29]

    X. Chen, J. Pan, J. Dong, Bidirectional multi-scale implicit neural representations for image deraining, in: IEEE Conf. Comput. Vis. Pattern Recog., 2024, pp. 25627–25636. 24

    2024

  22. [30]

    H. Chen, X. Chen, J. Lu, Y . Li, Rethinking multi-scale representations in deep deraining transformer, in: AAAI Conf. Artificial Intell., V ol. 38, 2024, pp. 1046–1053

    2024

  23. [31]

    ¨Ozdenizci, R

    O. ¨Ozdenizci, R. Legenstein, Restoring vision in adverse weather conditions with patch-based denoising di ffusion models, IEEE Trans. Pattern Anal. Mach. Intell. 45 (8) (2023) 10346–10357

    2023

  24. [32]

    J. Liu, Q. Wang, H. Fan, Y . Wang, Y . Tang, L. Qu, Residual denoising diffusion models, in: IEEE Conf. Comput. Vis. Pattern Recog., 2024, pp. 2773–2783

    2024

  25. [33]

    M. Wei, Y . Shen, Y . Wang, H. Xie, J. Qin, F. L. Wang, RainDiffusion: When unsupervised learning meets di ffusion models for real-world image deraining, arXiv e-prints (2023) arXiv–2301

    2023

  26. [34]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, arXiv preprint arXiv:2312.00752 (2023)

    2023 arXiv

  27. [35]

    T. Dao, A. Gu, Transformers are SSMs: Generalized models and e fficient algorithms through structured state space duality, in: Int. Conf. Mach. Learn., 2024

    2024

  28. [36]

    H. Li, Z. Liu, T. Zhao, T. Zhao, Y . Chen, N. Xie, MS-RainMamba: Learning multi-scale state space models for single image deraining, in: IEEE Int. Conf. Acoustics, Speech and Signal Process., IEEE, 2025, pp. 1–5

    2025

  29. [37]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: IEEE Conf. Comput. Vis. Pattern Recog., 2016, pp. 770–778

    2016

  30. [38]

    Sanghi, H

    A. Sanghi, H. Chu, J. G. Lambourne, Y . Wang, C.-Y . Cheng, M. Fumero, K. R. Malekshan, CLIP-Forge: Towards zero-shot text-to-shape generation, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 18603–18613

    2022

  31. [39]

    Y . Rao, W. Zhao, G. Chen, Y . Tang, Z. Zhu, G. Huang, J. Zhou, J. Lu, DenseCLIP: Language-guided dense prediction with context-aware prompting, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 18082–18091

    2022

  32. [40]

    Zhang, R

    R. Zhang, R. Fang, W. Zhang, P. Gao, K. Li, J. Dai, Y . Qiao, H. Li, Tip-Adapter: Training-free clip-adapter for better vision-language modeling, arXiv preprint arXiv:2111.03930 (2021)

    2021 arXiv

  33. [41]

    G. Kim, T. Kwon, J. C. Ye, DiffusionCLIP: Text-guided diffusion models for robust image manipulation, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 2426–2435

    2022

  34. [42]

    Y . Wang, L. Miao, Z. Zhou, L. Zhang, Y . Qiao, Infrared and visible image fusion with language-driven loss in clip embedding space, arXiv preprint arXiv:2402.16267 (2024)

    2024

  35. [43]

    Z. Luo, F. K. Gustafsson, Z. Zhao, J. Sj ¨olund, T. B. Sch¨on, Controlling vision-language models for universal image restoration, arXiv preprint arXiv:2310.01018 (2023)

    2023 arXiv

  36. [44]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional networks for biomedical image segmentation, in: N. Navab, J. Hornegger, W. M. Wells, A. F. Frangi (Eds.), Int. Conf. Med. Image Comp. Assist. Interv., Springer International Publishing, Cham, 2015, pp. 234–241

    2015

  37. [45]

    Ulyanov, A

    D. Ulyanov, A. Vedaldi, V . Lempitsky, Deep image prior, in: IEEE Conf. Comput. Vis. Pattern Recog., 2018, pp. 9446–9454

    2018

  38. [46]

    Lehtinen, J

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, T. Aila, Noise2Noise: Learning image restoration without clean data, in: J. Dy, A. Krause (Eds.), Int. Conf. Mach. Learn., V ol. 80 of Proceedings of Machine Learning Research, PMLR, 2018, pp. 2965–2974

    2018

  39. [47]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, in: International Conference on Learning Representations, 2019

    2019

  40. [48]

    D. Ren, W. Zuo, Q. Hu, P. Zhu, D. Meng, Progressive image deraining networks: A better and simpler baseline, in: IEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 3937–3946

    2019

  41. [49]

    Mehri, P

    A. Mehri, P. B. Ardakani, A. D. Sappa, MPRNet: Multi-path residual network for lightweight image super resolution, in: IEEE Winter Conf. Appl. Comput. Vis., 2021, pp. 2704–2713

    2021

  42. [50]

    B. Li, X. Liu, P. Hu, Z. Wu, J. Lv, X. Peng, All-in-one image restoration for unknown corruption, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 17452–17462

    2022

  43. [51]

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, Restormer: E fficient transformer for high-resolution image restoration, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 5728–5739

    2022

  44. [52]

    J. M. J. Valanarasu, R. Yasarla, V . M. Patel, Transweather: Transformer-based restoration of images degraded by adverse weather conditions, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 2353–2363. 25

    2022

  45. [53]

    Potlapalli, S

    V . Potlapalli, S. W. Zamir, S. H. Khan, F. Shahbaz Khan, Promptir: Prompting for all-in-one image restoration, Adv. Neural Inform. Process. Syst. 36 (2023) 71275–71293

    2023

  46. [54]

    Z. Yang, H. Chen, Z. Qian, Y . Yi, H. Zhang, D. Zhao, B. Wei, Y . Xu, All-in-one medical image restoration via task-adaptive routing, in: Int. Conf. Med. Image Comp. Assist. Interv., Springer, 2024, pp. 67–77

    2024

  47. [55]

    OpenAI, GPT-4 technical report, arXiv preprint arXiv:2303.08774 (2024). 26

    2024 arXiv

Pith tools

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