DnCNN-tensorflow

DnCNN implement based on tensorflow-1.8

Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

To run this code

• Python3 with dependencies: scipy numpy tensorflow-gpu scikit-image pillow h5py

Generating training data

• 'generate_data.py'. You may need to modify the path to trainning datasets. According to the MatConvNet code offered by authors, there are some blank(zero) data in the generated training dataset.

Trainning

• [train] 'train_DnCNN.py'

Testing

• [test] 'Validate_DnCNN.py' You need to change the path and filetype of your testset. While I generated the testset by matlab to make faire when competing with other methods.

Results

• Only denoising methods were focused in my work.

Gaussian Denoising

The average PSNR(dB) results of different methods on the BSD68 dataset.

Noise Level	BM3D	WNNM	EPLL	MLP	CSF	TNRD	DnCNN-S	DnCNN-B	DnCNN-S-Re
15	31.07	31.37	31.21	-	31.24	31.42	31.73	31.61	-
25	28.57	28.83	28.68	28.96	28.74	28.92	29.23	29.16	-
50	25.62	25.87	25.67	26.03	-	25.97	26.23	26.23	-

The average PSNR(dB) results of different methods on the Set12 dataset.

Noise Level	BM3D	WNNM	EPLL	MLP	CSF	TNRD	DnCNN-S	DnCNN-B	DnCNN-S-Re
15	32.372	32.696	32.138	-	32.318	32.502	32.859	32.680	-
25	29.969	30.257	29.692	30.027	29.837	30.055	30.436	30.362	30.33
50	26.722	27.052	26.471	26.783	-	26.812	27.178	27.206	-