• NBNet: Noise Basis Learning for Image Denoising with Subspace Projection
• • Dependencies
  • Training
  • • Preparation
    • Begin training:
    • Begin testing:
  • Pretrained model

NBNet: Noise Basis Learning for Image Denoising with Subspace Projection

Code for CVPR21 paper NBNet.

The illustration of our key insight:

Dependencies

• MegEngine >= 1.3.1 (For DistributedDataParallel)

Training

Preparation

python prepare_data.py --data_dir yours_sidd_data_path

Begin training:

For SIDD benchmark, use:

python train_mge.py -d prepared_data_path -n num_gpus

For DnD benchmark, we use MixUp additionally:

python train_mge.py -d prepared_data_path -n num_gpus --dnd

Begin testing:

Download the pretrained checkpoint and use:

python test.py -d prepared_data_path -c checkpoint_path

The result is PSNR 39.765.

Pretrained model

MegEngine checkpoint for SIDD benchmark can be downloaded via
Google Drive
or
GitHub Release.