Guiding WaveMamba with Frequency Maps for Image Debanding

Official Code for the following paper:

X. Wang, S.Tasmoc, N. Anantrasirichai, and A. Katsenou. Guiding WaveMamba with Frequency Maps for Image Debanding

This paper was accepted by the Picture Coding Symposium. (PCS 2025).

---

Performance

Quantitative evaluation on the deepDeband test set (patch pairs)

Model	PSNR↑	SSIM↑	LPIPS↓	CAMBI↓	BBAND↓	DBI↓
Banded Images	36.286	0.977	0.081	1.467	1.217	0.083
FFmpeg	31.383	0.941	0.110	0.955	0.632	0.035
PPP	32.432	0.927	0.188	3.619	1.230	0.005
DIP	34.364	0.957	0.093	1.728	0.820	0.034
LDM	30.597	0.905	0.140	0.824	0.853	0.057
deepDeband-F	34.584	0.964	0.071	0.057	0.464	0.122
deepDeband-F-WWM	35.005	0.967	0.066	0.055	0.484	0.104
WaveMamba	42.399	0.989	0.037	3.947	1.410	0.021
WaveMamba-WWM	42.201	0.988	0.033	3.615	1.436	0.021
WaveMamba-DWT	42.111	0.988	0.036	3.798	1.403	0.022
WaveMamba-MAP	41.546	0.988	0.037	3.523	1.343	0.022

Quantitative evaluation on the HD_Images_DBI dataset (Full HD pairs)

Which serves as the source for patch extraction in the deepDeband dataset.

Method	PSNR↑	SSIM↑	LPIPS↓	CAMBI↓	BBAND↓	DBI↓
Banded Images	35.632	0.958	0.066	0.337	0.411	0.312
deepDeband-F	32.956	0.906	0.079	0.375	0.154	0.166
deepDeband-F-WWM	33.237	0.907	0.078	0.374	0.153	0.168
WaveMamba	39.037	0.977	0.046	0.543	0.386	0.133
WaveMamba-WWM	39.339	0.977	0.044	0.525	0.398	0.139
WaveMamba-DWT	38.97	0.977	0.045	0.616	0.412	0.128
WaveMamba-MAP	38.826	0.977	0.048	0.519	0.365	0.082

Cross-dataset evaluation on BAND-2k dataset, using models trained on deepDeband

Model	PSNR↑	SSIM↑	LPIPS↓	CAMBI↓	BBAND↓	DBI↓
Banded Images	37.344	0.964	0.100	0.513	0.564	0.431
deepDeband-F	33.130	0.897	0.116	0.561	0.198	0.225
deepDeband-F-WWM	33.263	0.898	0.115	0.557	0.198	0.227
WaveMamba	38.926	0.977	0.063	0.647	0.519	0.148
WaveMamba-WWM	39.054	0.977	0.059	0.639	0.541	0.157
WaveMamba-DWT	38.963	0.977	0.062	0.702	0.551	0.158
WaveMamba-MAP	38.505	0.976	0.066	0.691	0.476	0.082

Best and second-best results are highlighted in bold and italic underline, respectively. Arrows indicate if higher or lower values are preferred.

See stat_metric.ipynb for a comprehensive performance comparison and detailed statistical results.

Visual Inspection

Visual comparison of debanding performance during inference on the BAND-2k dataset

Qualitative comparisons of the models trained and tested on deepDeband patches

Full HD visual comparison of debanding results on a difficult case

More visual examples can be found in the test_img folder.

Banded	deepDeband-f	deepDeband-f-WWM	WaveMamba
WaveMamba-WWM	WaveMamba-DWT	WaveMamba-MAP	Pristine

Usage

📥 Download Banding Datasets

The corresponding bading image datasets can be downloaded from the following sources:
deepDeband, HD_images_dataset_dbi, BAND-2k.

🔧 WaveMamba with Frequency Maps

To implement the following variants, please refer to the README.md in WaveMamba-Frequency-Masking:

1. WaveMamba-WWM, 2. WaveMamba-DWT, 3. WaveMamba-MAP

📈 Evaluation Metrics

We evaluate performance using the following metrics: PSNR, SSIM, LPIPS, CAMBI, BBAND, DBI.

Run the following script to compute quality metrics and statistical results:

python src/cal_metrics.py
python src/cal_stat.py

We include computed BBAND and DBI scores are included in: src/BBAND/ and src/DBI/. For details on DBI metric training and computation, please refer to the original DBI GitHub repository.

Acknowledgements

Based on BasicSR and Wave-Mamba framework.

Citation

If you find this paper and the repo useful, please cite our paper 😊:

@article{wang2025wavemamba-frequency-map,
  title={Guiding WaveMamba with Frequency Maps for Image Debanding},
  author={Wang, Xinyi and Tasmoc, Smaranda and Anantrasirichai, Nantheera and Katsenou, Angeliki},
  booktitle={Picture Coding Symposium (PCS 2025)},
  year={2025},
}

Contact:

Xinyi WANG, xinyi.wang@bristol.ac.uk