JiaHongZ / CDN

CDN

This code is a implement of the paper "Considering Image Information and Self-similarity: A Compositional Denoising Network".

Datasets:

Trainsets

• For Gaussian denoising on gray images and color images

  DIV2K

• For real noise

  Trainsets from SIDD

Testsets

• For gray images

set12

bsd68

• For color images

set5

Kodak24

• For real noise

Validation data from SIDD(http://130.63.97.225/sidd/dataset.php)

DND(https://noise.visinf.tu-darmstadt.de/)

Train

main_train_###.py

The file main_train_compdnet_###.py contains the best model in the paper:

• For Gaussian denoising on gray images: main_train_compdnet.py
• For Gaussian denoising on color images : main_train_compdnet_color.py
• For real noise: main_train_real.py

Testing

main_test_###.py

For Gaussian noise, we provide pre-trained models of three noise level, 15, 25 and 50. Please ensure the test noise level is consistent with the model you select (you can select them in the test file).

For real noise, we provide the pre-trained model. Test results can be produced by running file as follows:

test_denoising_sidd.py for SIDD valuation.

test_denoising_dnd.py for DND. The file generated by it also should be submitted.

Pre-trained model:

Link: https://pan.baidu.com/s/1EOUBHBe4p9v4yQIYc2IV3Q?pwd=nviv Code: nviv

Download and put them in the model_zoo file.

Directory structure

• data (data processing and data loaders)
• models (network, train method definition)
• model_zoo (pre-trained models)
• testsets
• trainsets
• utils