Skip to content
No description, website, or topics provided.
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
data/triangle
README.md
dip.py
make_dataset.py
relunet.py
skip.py
utils.py

README.md

The Spectral Bias of the Deep Image Prior

Code to reproduce results in the paper: https://arxiv.org/abs/1912.08905

Dataset

Download images from Table 2 here: http://www.cs.tut.fi/~foi/GCF-BM3D/index.html. Save these images in a folder called data/

Generate noisy images:

python make_dataset.py --downsample_factors 1,2,4

This will add different levels of noise to the images downsampled by the specified factors and save them in the data/ folder

Trajectory Experiment

Run the following to reproduce the trajectory experiment (figure 1 in the paper)

# Less high frequency components
python dip.py --noisy_img data/triangle/triangle-0.2.png --niter 1000 --traj_iter 10

# More high frequency components
python dip.py --noisy_img data/triangle/triangle-0.9.png --niter 1000 --traj_iter 10

Denoising Experiment

Run the following to reproduce the denoising results with different architectures (tables 1 and 2 in the paper):

  1. DIP model (a convolutional encoder-decoder):

     python dip.py --noisy_img data/denoise-4/House256/House256_s25.png --clean_img data/denoise-4/House256/House256.png
    
  2. ReLUNet:

     python relunet.py --noisy_img data/denoise-4/House256/House256_s25.png --clean_img data/denoise-4/House256/House256.png
    
  3. DIP Linear-128:

     python dip.py --linear --n_ch_up 128 --n_ch_down 128 --noisy_img data/denoise-4/House256/House256_s25.png --clean_img data/denoise-4/House256/House256.png
    
  4. DIP Linear-2048:

     python dip.py --linear --n_ch_up 2048 --n_ch_down 2048 --noisy_img data/denoise-4/House256/House256_s25.png --clean_img data/denoise-4/House256/House256.png
    

This runs DIP twice to generate 2 trajectories. The denoising results, along with variation of distance between the 2 trajectories will be saved in the outputs/ folder.

You can’t perform that action at this time.