Image deraining aims to recover rain-free images from rain-degraded observations. Recent frequency-aware deraining methods have demonstrated promising performance by removing rain artifacts in the spectral domain. However, since rain artifacts and background textures often appear in several correlated subbands, prior frequency-domain deraining models still enhance frequency subbands separately. Such band-wise processing limits information exchange across subbands. Without context from other subbands, rain and clean content that appear together in one subband cannot be separated. This leads to either residual rain streaks or over-smoothing of fine details. To address this limitation, we propose AFRNet, an Adaptive Frequency Redistribution Network for image deraining. First, we introduce an Adaptive Frequency Recombination Module (AFRM) that performs cross-band modulation among wavelet subbands, enabling information interaction and adaptive recombination of degraded frequency components. We then develop a Subband Compensation Block (SCB) that compensates for decoder subband degradation by leveraging encoder subband features through inter-subband attention in the wavelet domain, alleviating detail loss during reconstruction. Extensive experiments demonstrate that AFRNet achieves competitive performance and consistent results across diverse deraining benchmarks.
Figures & tables
Fig. 1: Band-wise processing vs. cross-band redistribution. Band-wise processing retains residual artifacts, while cross-band redistribution enables cleaner restoration. Insets show subband interaction patterns (diagonal: isolated processing; off-diagonal: cross-band exchange).
Fig. 2: Frequency analysis of rain degradation. Rain streaks and raindrops show different frequency properties: rain streaks are more anisotropic, whereas raindrops are more isotropic. Single-band processing cannot handle both types well, which motivates cross-band modulation.
Fig. 3: Architecture of AFRNet for image deraining. AFRNet combines frequency redistribution with spatial feature restoration. Its main components are: (a) the Adaptive Frequency Recombination Module (AFRM), which performs content-adaptive cross-subband modulation to redistribute frequency information; (b) the Subband Compensation Block (SCB), which transfers encoder subband features to decoder subbands to recover details lost during reconstruction; and (c) the Adaptive Frequency Restoration Block (AFRBlock), which combines local spatial processing with standard Fourier mixing to refine spatial features. The wavelet encoder-decoder provides the feature hierarchy and subband interfaces for these operations. WT Down and IWT Up use the wavelet transform and inverse wavelet transform to replace convolution-based downsampling and upsampling.
Method
Year
Test100
Rain100H
Rain100L
Test2800
Test1200
Average
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
RESCAN [ 10 ]
ECCV2018
21.59
0.726
18.01
0.467
24.15
0.791
24.50
0.765
24.40
0.759
22.53
0.702
PReNet [ 8 ]
CVPR2019
23.17
0.752
17.63
0.487
27.76
0.876
27.20
0.825
26.05
0.792
24.36
0.746
SPDNet [ 44 ]
ICCV2021
24.25
0.848
25.87
0.809
28.63
0.880
31.05
0.904
30.42
0.893
28.04
0.867
PCNet [ 45 ]
TIP2021
23.29
0.762
20.83
0.563
26.64
0.817
27.10
0.818
26.53
0.791
24.88
0.750
MPRNet [ 19 ]
CVPR2021
25.66
0.859
28.23
0.850
31.94
0.930
32.14
0.925
31.32
0.901
29.86
0.893
TABLE I: Quantitative Comparison on the Five Rain13K Test Sets in Terms of Y-Channel PSNR and SSIM. Best and second-best results are shown in bold and underlined , respectively.
Dataset
MPRNet [ 19 ]
Uformer [ 13 ]
Restormer [ 11 ]
DRSformer [ 21 ]
MambaIR [ 59 ]
DFSSM [ 33 ]
AFRNet-S
AFRNet-L
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
Rain200L
39.47
0.983
40.20
0.986
40.99
0.989
41.23
0.989
41.13
0.990
41.81
0.991
42.28
0.990
42.52
0.991
Rain200H
30.67
0.911
30.80
0.911
32.00
0.933
32.17
0.933
32.18
0.930
32.99
0.940
32.90
0.940
33.02
0.944
SPA-Data
43.64
0.984
46.13
0.991
47.98
0.992
48.54
0.992
–
–
49.55
0.994
49.83
0.994
49.90
0.995
TABLE II: Quantitative Comparison on Rain200L, Rain200H, and SPA-Data in Terms of Y-Channel PSNR and SSIM. “–” denotes an unavailable result. Best and second-best results are shown in bold and underlined , respectively.
Method
Venue
RS
RD
RDRS
MPRNet [ 19 ]
CVPR2021
40.81/0.981
37.03/0.972
34.79/0.956
DGUNet [ 60 ]
CVPR2022
41.09/0.983
37.56/0.975
35.34/0.959
Uformer [ 13 ]
CVPR2022
40.69/0.972
37.08/0.966
34.99/0.954
Restormer [ 11 ]
CVPR2022 (Oral)
41.42/0.980
38.78/0.976
36.08/0.961
IDT [ 61 ]
CVPR2022
41.61/0.983
39.09/0.980
36.23/0.960
NAFNet [ 49 ]
ECCV2022
40.39/0.972
37.23/0.974
34.99/0.957
TABLE III: Quantitative Comparison on the Official RainDS-Syn 3,000/600 Split. Each entry presents Y-channel PSNR/SSIM; the three columns correspond to streak (RS), raindrop (RD), and compound-rain (RDRS) degradation.
Fig. 4: Qualitative comparison on Rain100L. Enlarged regions highlight residual streak suppression and fine-detail preservation.
Fig. 5: Qualitative comparison on Rain100H under heavy-rain conditions. Enlarged regions highlight streak removal and structure preservation.
Fig. 6: Qualitative comparison on SPA-Data. Enlarged regions highlight streak suppression and edge preservation in real rainy scenes.
Fig. 7: Qualitative results on RainDS-Syn under streak, raindrop, and compound-rain degradation types. Residual maps visualize the removed rain components.
Fig. 8: Qualitative results on unpaired RainDS-Real images.
Method
PSNR ↑
SSIM ↑
DMPHN [ 63 ]
31.20
0.940
Restormer [ 11 ]
32.92
0.961
SFNet [ 64 ]
33.27
0.963
ACL [ 65 ]
33.25
0.964
AFRNet (Ours)
33.57
0.968
TABLE IV: Quantitative comparison on GoPro image deblurring in terms of RGB PSNR and SSIM. The best results are shown in bold.
Fig. 9: Qualitative comparison on GoPro image deblurring.
Method
Params (M) ↓
GFLOPs ↓
DRSformer [ 21 ]
33.7
242.9
TransWeather [ 26 ]
31.0
–
AdaIR [ 30 ]
28.77
147.45
UDR-S2Former [ 22 ]
28.53
21.58
Restormer [ 11 ]
26.12
174.7
NeRD-Rain [ 23 ]
22.89
156.3
TABLE V: Model Complexity for a 256×256 Input. “–” denotes an unavailable profile.
Fig. 10: Visualization of the wavelet-domain operation of AFRM. (a) Input rainy image. (b) Wavelet subbands of the input image. (c) Content-adaptive modulation by AFRM. (d) Modulated wavelet subbands. (e) Image reconstructed from the modulated subbands.
Model
AFRM
SCB
PSNR ↑
SSIM ↑
Backbone (baseline)
×
×
31.25
0.928
+ AFRM (within-band modulation)
within-band
×
31.38
0.930
+ AFRM (full cross-band mixing)
✓
×
32.01
0.934
+ SCB (independent β )
×
independent β
31.49
0.931
+ SCB (full compensation)
×
✓
32.14
0.935
AFRNet (Full)
✓
✓
33.02
0.944
TABLE VI: Ablation Study on Rain200H.
Fig. 11: Visualization of SCB compensation across wavelet subbands. The four rows correspond to the LL , LH , HL , and HH subbands, respectively. From left to right, the columns show the subband before SCB, the compensation produced by SCB, and the subband after SCB.
While deep learning has advanced single-image deraining, existing models suffer from a fundamental limitation: they employ a static inference paradigm that fails to adapt to the complex, coupled degradations (e.g., noise artifacts, blur, and color deviation) of real-world rain. Consequently, restored images often exhibit residual artifacts and inconsistent perceptual quality. In this work, we present Derain-Agent, a plug-and-play refinement framework that transitions deraining from static processing to dynamic, agent-based restoration. Derain-Agent equips a base deraining model with two core capabilities: 1) a Planning Network that intelligently schedules an optimal sequence of restoration tools for each instance, and 2) a Strength Modulation mechanism that applies these tools with spatially adaptive intensity. This design enables precise, region-specific correction of residual errors without the prohibitive cost of iterative search. Our method demonstrates strong generalization, consistently boosting the performance of state-of-the-art deraining models on both synthetic and real-world benchmarks.
Despite significant progress has been made in image deraining, we note that most existing methods are often developed for only specific types of rain degradation and fail to generalize across diverse real-world rainy scenes. How to effectively model different rain degradations within a universal framework is important for real-world image deraining. In this paper, we propose UniRain, an effective unified image deraining framework capable of restoring images degraded by rain streak and raindrop under both daytime and nighttime conditions. To better enhance unified model generalization, we construct an intelligent retrieval augmented generation (RAG)-based dataset distillation pipeline that selects high-quality training samples from all public deraining datasets for better mixed training. Furthermore, we incorporate a simple yet effective multi-objective reweighted optimization strategy into the asymmetric mixture-of-experts (MoE) architecture to facilitate consistent performance and improve robustness across diverse scenes. Extensive experiments show that our framework performs favorably against the state-of-the-art models on our proposed benchmarks and multiple public datasets.
Qianfeng Yang, Qiyuan Guan, Xiang Chen +3
Dalian Polytechnic University · Nanjing University of Science and Technology
Blind all-in-one image restoration must adapt a shared model to heterogeneous and sometimes mixed degradations without a task label at test time. We introduce WaFIR, a decoder module that decomposes features into Haar subbands, summarizes each subband with a compact Fourier descriptor, and routes it to prompt-conditioned spectral modulation experts. Unlike image-level frequency conditioning, WaFIR makes routing decisions at the subband level, enabling frequency-specific processing while retaining the task context learned by the backbone. Under a unified training and evaluation protocol for denoising, deraining, and dehazing, WaFIR improves average PSNR over the three tasks by 0.48 dB over the PromptIR baseline and by 0.38 dB over the recent Up-Restorer. Matched-capacity and routing controls show that the gain is not explained by additional parameters alone. On unseen noise intensities, descriptor-based routing is more robust than a discrete label-conditioned alternative and remains on par under compound degradations, while routing analyses show stable subband-dependent expert use across degradation types. These results support subband frequency structure as an interpretable complement to prompt-based conditioning for blind image restoration. Code is available at https://anonymous.4open.science/r/WaFIR/README.md
Trong Nhan Nguyen, Vu Hoang Minh Doan, Thanh Dat Le +6
Pukyong National University, Busan, Republic of Korea · University of Alabama at Birmingham, Birmingham, AL, USA