Keras implementation of DnCNN-S. Originaly as proposed by Zhang et al in the paper Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. This implementation is only for DnCNN-S (Specified noise level).
-
This project is useful to denoise an image, if the noise level in an image is known or estimated. You just need to change the scale parameter in the file conf/myConfig.py.
-
This project is also useful to know "How to create custom loss, custom real time data-augmentation flow and custom learning rate scheduler in keras?", check kDnCNN.py for that, I had a hard time figuring it out myself.
- python 3, keras 2(tf-backend), OpenCV 3 were being used for development.
$ python generateData.py #this will create new folder name trainingPatch containg image patches.
$ python kDnCNN.py #to train, and it saves model myModel.h5 in your working directory.
$ python testPSNR.py --dataPath /path/to/test/dataset/ --weightsPath /path/to/myModel.h5 #to calculate avg PSNR on test data
- I have used Set12/ dataset for testing. Avg PSNR(dB)
| Noise Level | DnCNN-S | KDnCNN-S |
|---|---|---|
| 25 | 30.4 | 28.3 |
- NOTE: Any suggestion to improve performance of KDnCNN-S to make it at par with original DnCNN-S is welcomed.