Skip to content

Latest commit

 

History

History
46 lines (30 loc) · 829 Bytes

README.md

File metadata and controls

46 lines (30 loc) · 829 Bytes

DnCNN-keras

A keras implemention of the paper Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

Dependence

tensorflow
keras2
numpy
opencv

Prepare train data

$ python data.py

Clean patches are extracted from 'data/Train400' and saved in 'data/npy_data'.

Train

$ python main.py

Trained models are saved in 'snapshot'.

Test

$ python main.py --only_test True --pretrain 'path of saved model'

Noisy and denoised images are saved in 'snapshot'.

Results

Gaussian Denoising

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

Noise Level BM3D DnCNN-S DnCNN-keras
25 28.57 29.23 29.21