Skip to content

Latest commit

 

History

History
105 lines (69 loc) · 3.67 KB

tunable_conv.md

File metadata and controls

105 lines (69 loc) · 3.67 KB
Raw

NAFNet for Modulated Image Denoising

  • Dataset

    Need to prepare SIDD Srgb validation dataset by running.

  • Inference Test Case

    Then run inference as

    python test_tunable_nafnet.py

  • Tunable Parameters

    The tunable parameters should always sum to one. In this testcase, we control the denoising strength. For example 1.0 0.0 will provide maximum denoising, 0.5 0.5 will provide medium denoising, 0.0 1.0 will provide minimum denoising.

    For visualization, you can set flag to save image. The inference images will be saved in output/nafnet, and each png will contain (from left to right): noisy input, predicted output, ground-truth.

SwinIR for Modulated Image Denoising and Perceptual Super-Resolution

  • Dataset

    Need to prepare Kodak dataset.

  • Inference Test Case

    Main interface is very similar as before, just run

    python test_tunable_nafnet.py

    and by default the denoising experiment with standard deviation 25 will start.

  • Tunable Parameters

    As before, you can change tunable parameters, add flag to save image, and also add reference PyTorch implementation. In addition to that you can also select which experiment to run.

    Where noise_stddev will run the denoising experiment (Gaussian noise with standard deviation 15/25/50, and sr_factor the super-resolution experiment (super-resolution factor 4).

    In both experiments, we have two tunable parameters. In the denoising experiment, the parameters control the denoising strength. For example 1.0 0.0 will provide maximum denoising, and 0.0 1.0 will provide minimum denoising. In the super-resolution experiment the parameters control the perception-distortion tradeoff. For example 1.0 0.0 will provide maximum accuracy, and 0.0 1.0 will provide maximum perceptual quality.

EDSR for Modulated Joint Image Denoising and Deblurring

  • Dataset

    Need to prepare Kodak dataset.

  • Inference Test Case

    Main interface is very similar as before, just run

    python test_tunable_edsr.py

  • Tunable Parameters

    As before, you can change tunable parameters, add flag to save image, and also add reference PyTorch implementation. In addition to that you can also select the noise standard deviation as well as the blur size via the following arguments:

    Valid values of noise_stddev are in [5, 30], and for blur_stddev are in [0,4]. Low noise_stddev values correspond to low amount of noise in the input image, and low blur_stddev values correspond to less blur in the input image.

    Note that in this case we again have two tunable parameters: the first parameter control the denoising strength, and the second the deblurring strength. For example 1.0 0.0 will provide maximum denoising and minimum deblurring, and 1.0 1.0 will provide maximum denoising and deblurring.

StyleNet for Modulated Style Transfer

  • Dataset

    Need to prepare Kodak dataset.

  • Inference Test Case

    Main interface is very similar as before, just run

    python test_tunable_edsr.py

  • Tunable Parameters

    As before, you can change tunable parameters using --params, add flag to save image --save_images. Note that in this case we again three tunable parameters controlling the influence of three different styles, namely, Mosaic (1.0 0.0 0.0), Edtaonisl (0.0 1.0 0.0), and Kandinsky (0.0 0.0 1.0). You can try any combination of parameters, as long as the sum of all parameters is 1.0; for instance 0.0 0.5 0.5 is correct, but 1.0 0.5 0.0 is not.