Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1037006(2025)
Two-Stage Image Rain Removal Network Based on Dynamic Residual Diffusion
Figures & Tables(13)
Fig. 1. Overall flowchart of proposed method. (a) Rain removal stage; (b) detail recovery stage
Fig. 5. Pseudocode for training and inference processes. (a) Training process; (b) inference process
Fig. 6. Detailed comparisons of rain removal effect of different methods tested on Rain100H dataset. (a) Rainy images; (b) JORDER; (c) RESCAN; (d) PReNet; (e) MPRNet; (f) DRSformer; (g) MFDNet; (h) ours; (i) ground truth
Fig. 7. Detailed comparisons of rain removal effect of different methods tested on real dataset. (a) Real rainy images; (b) DDN; (c) JORDER; (d) PReNet; (e) Syn2Real; (f) DRSformer; (g) MFDNet; (h) ours
Fig. 8. Detailed comparisons of rain removal effect of different methods tested on real dataset. (a) Real rainy images; (b) DDN; (c) JORDER; (d) PReNet; (e) Syn2Real; (f) DRSformer; (g) MFDNet; (h) ours
Fig. 9. Detailed comparisons of rain removal effect of different module ablation experiments. (a) Rain image; (b) real image; (c) baseline; (d) baseline+DAB; (e) baseline+DAB+cascade repair module; (f) baseline+DAB+ residual diffusion repair module; (g) visual image of repaired structure
Table 1. PSNR and SSIM metrics of different algorithms on synthetic datasets
View tableTable 1. PSNR and SSIM metrics of different algorithms on synthetic datasets
Method Rain100H Rain100L Rain800 Rain1200 PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM PSNR/dB SSIM DDN［ 5 ］21.94 0.768 32.15 0.931 24.04 0.867 27.28 0.871 JORDER［ 25 ］26.54 0.835 36.11 0.970 23.47 0.869 24.33 0.861 RESCAN［ 31 ］26.36 0.786 36.21 0.975 25.00 0.835 30.51 0.882 PReNet［ 9 ］29.46 0.899 37.48 0.979 24.81 0.851 31.36 0.911 DRD-Net［ 20 ］24.58 0.863 33.27 0.956 26.32 0.901 26.73 0.824 MPRNet［ 7 ］30.41 0.889 36.40 0.965 30.27 0.897 32.91 0.916 Restormer［ 14 ］31.46 0.904 38.99 0.978 32.00 0.923 32.99 0.926 DRSformer［ 17 ］32.03 0.903 39.05 0.980 31.71 0.911 33.62 0.928 MFDNet［ 32 ］30.48 0.899 37.61 0.973 30.78 0.914 33.01 0.925 Ours 31.75 0.912 39.12 0.981 32.21 0.925 33.07 0.930
Table 2. NIQE and SSEQ metrics of different algorithms on real dataset
View tableTable 2. NIQE and SSEQ metrics of different algorithms on real dataset
Method NIQE SSEQ DDN［ 5 ］3.81 26.74 JORDER［ 25 ］3.42 24.37 RESCAN［ 31 ］4.02 26.96 PReNet［ 9 ］4.24 27.15 MPRNet［ 7 ］3.89 26.88 Syn2Real［ 33 ］4.31 27.64 DRSformer［ 17 ］3.43 24.59 MFDNet［ 32 ］3.57 25.30 Ours 3.32 23.89
Table 3. Ablation experimental results of different modules
View tableTable 3. Ablation experimental results of different modules
Method Model structure Rain100H DAB Cascading recovery module Residual diffusion recovery module PSNR /dB SSIM 1 — — — 28.32 0.887 2 √ — — 29.73 0.895 3 √ √ — 29.94 0.903 4 √ — √ 31.75 0.912
Table 4. Results of ablation experiments with different sampling steps
View tableTable 4. Results of ablation experiments with different sampling steps
Sampling step PSNR /dB SSIM 2 29.73 0.895 3 31.75 0.912 4 31.44 0.908 5 30.97 0.897 10 29.50 0.872
Tools
Get Citation
Copy Citation Text
Yixin Yang, Xinjian Gao, Ye Ma, Jun Gao. Two-Stage Image Rain Removal Network Based on Dynamic Residual Diffusion[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1037006
Paper Information
Category: Digital Image Processing
Received: Oct. 15, 2024
Accepted: Nov. 26, 2024
Published Online: Apr. 23, 2025
The Author Email:
CSTR:32186.14.LOP242104
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20