Skip to content

Latest commit

 

History

History
107 lines (68 loc) · 5.45 KB

Synthetic.md

File metadata and controls

107 lines (68 loc) · 5.45 KB

Prerequisites

  1. Refer to the INSTALL.md for instructions on preparing environment and dependencies.

  2. Download training and testing datasets from One Drive.

Train

  1. For training, simply place the training datasets in './datasets/fivek/Raw'. We randomly select 220 relatively clean raw images from it: 200 for training and 20 for testing. The small testing set is used for fast inference.

    After placing the files, ensure the following structure:

    datasets/
├── fivek/
│   └── list_file/
│       ├── invalid_list.txt
│       ├── train_list.txt
│       ├── val_list.txt
│   └── Raw/
│       ├── a0001-jmac_DSC1459.dng
│       ├── a0002-dgw_005.dng
│       ├── ...
│       ├── a5000-kme_0204.dng

  2. Run

    • To train our DuanDn model, run 

      python train_dualdn.py -opt ./options/DualDn.yml

    • To train the raw-denoising model with the same backbone, run 

      python train_dualdn.py -opt ./options/Raw_Denoising.yml

    • To train the sRGB-denoising model with the same backbone, run 

      python train_dualdn.py -opt ./options/sRGB_Denoising.yml

    • You can adjust the default training settings in the datasets/train section of the ./options/****.yml file, including but not limited to parameters such as the random noise level (K_min, K_max), fixed noise type (noise_model), random ISP amplification ratio (alpha__min, alpha__max) and different ISP algorithms (final_stage, demosaic_type, gamma_type).

    • You can adjust the default denoising paradigms in the network section of the ./options/****.yml file. This includes parameters such as the denoising domain (type), denoising backbone (backbone_type), channel number (c) and noise map usage (with_noise_map).

  3. For validation, we evaluate 20 synthetic images and calculating 4 evaluation metrics, that is PSNR, SSIM, NIQE and LPIPS.

    • The random seed is fixed, ensuring that the generated synthetic noisy raw remains consistent in each validation round.
    • For faster validation, open the ./options/****.yml file and set central_crop in datasets/val to true. This setting only validates the central square crop of each image, with a size of (patch_size, patch_size). Afterward, you can choose to test the full image with the following test code.
    • By default, validation uses noise levels [0.002, 0.02] and ISP amplification ratio [0, 0.5], creating a total of 4 validation sets (2 noise levels * 2 ratios). You can modify these settings in the datasets/val section of the ./options/****.yml file for custom validation configurations.

  4. Find the training results in './experiments'

Test

  1. After training, you can test the trained model in various testing sets, here we test Synthetic images for example.

  2. Run

    python test_dualdn.py -opt [exp_option_path] --num_iters [iters] --val_datasets ['Synthetic', 'Real_captured', 'DND']

    E.g. If you trained DualDn with 'DualDn.yml' for 120000 iterations, and want to test it on Synthetic datasets:

    python test_dualdn.py -opt ./experiments/DualDn/DualDn.yml --num_iters 120000 --val_datasets Synthetic

  3. You can modify the test settings in the datasets/val section of the experiment options file at './experiments/DualDn/DualDn.yml', including but not limited to parameters such as the random noise level (K_min, K_max), fixed noise type (noise_model), random ISP amplification ratio (alpha__min, alpha__max) and different ISP algorithms (final_stage, demosaic_type, gamma_type).

  4. Find the testing results in './results'

Inference

  1. For fast inference, you can use the pre-trained models to process specific images with selected synthetic noise level and ISP amplification ratio.

  2. Download the pre-trained model from One Drive and place it in './pretrained_model'

  3. According to the raw file in './datasets/fivek/Raw', add the correct filenames to './datasets/fivek/list_file/val_list.txt', with one filename per line.

  4. Run

    python inference_dualdn.py -opt [exp_option_path] --pretrained_model [model_path] --noise_level [[noise1, noise2...]] --alpha [[alpha1, alpha2...]]

    E.g. If you download the pre-trained model of DualDn_Restormer.pth, and want to test it with noise levels [0.002, 0.02] and ISP amplification ratio [0, 0.5]:

    python inference_dualdn.py -opt ./options/DualDn.yml --pretrained_model ./pretrained_model/DualDn_Restormer.pth --noise_level [0.002, 0.02] --alpha [0, 0.5]

  5. Find the inferencing results in './results'