Chadkowski/ADL

ADL: Adversarial Distortion Learning for Denoising and Distortion Removal

Morteza Ghahremani, Mohammad Khateri, Alejandra Sierra, Jussi Tohka

AiVi, UEF, Finland

---

This repository is the official implementation of ADL: Adversarial Distortion Learning for denoising medical and computer vision images (arxiv, supp, pretrained models, visual results).

TensorFlow	PyTorch

ADL achieves state-of-the-art Gaussian denoising performance in

• grayscale/color image denoising in Medical imaging 🔥🔥🔥
• grayscale/color image denoising in Computer Vision images 🔥🔥🔥
• JPEG compression artifact reduction 🔥🔥🔥
• grayscale/color deblurring 🔥🔥🔥

Network architectures

• Proposed Efficient-UNet (Denoiser)

• Proposed Efficient-UNet (Discriminator)

______________

Denoising Results on BSD68 and CBSD68:

• Results on the BSD68 dataset for Additive white Gaussian noise:

σ	BM3D	WNNM	DnCNN	NLRN	FOCNet	MWCNN	DRUNet	SwinIR	ADL (ours)
15	31.08	31.37	31.73	31.88	31.83	31.86	31.91	31.97	🔥 32.11 🔥
25	28.57	28.83	29.23	29.41	29.38	29.41	29.48	29.50	🔥 29.50 🔥
50	25.60	25.87	26.23	26.47	26.50	26.53	26.59	26.58	🔥 26.87 🔥

• Here we reported the results of the techniques reported by the authors.
• Our ADL was trained on the grey Flickr2K dataset only!

CBSD68 (img_id: test015)	Noisy (σ=25)	SwinIR	ADL (ours)

• Results on the CBSD68 dataset for Additive white Gaussian noise:

σ	BM3D	WNNM	EPLL	MLP	CSF	TNRD	DnCNN	DRUNet	SwinIR	ADL (ours)
15	33.52	33.90	33.86	33.87	33.91	-	34.10	34.30	34.42	🔥 34.61 🔥
25	30.71	31.24	31.16	31.21	31.28	31.24	31.43	31.69	31.78	🔥 32.18 🔥
50	27.38	27.95	27.86	27.96	28.05	28.06	28.16	28.51	28.56	🔥 29.02 🔥

CBSD68 (img_id: test015)	Noisy (σ=50)	SwinIR	ADL (ours)

Denoising Results on Medical Images:

2D (click here)

3D MRI Brain-BrainWeb (click here)

3D MRI knee-fastMRI (click here)

---

Citation

If you find ADL useful in your research, please cite our tech report:

@article{ADL2022,
    author = {Morteza Ghahremani, Mohammad Khateri, Alejandra Sierra, Jussi Tohka},
    title = {Adversarial Distortion Learning for Medical Image Denoising},
    journal = {arXiv:2204.14100},
    year = {2022},
}