LightField Denoising

This repository contains the code for the student research project titled "Light field denoising using conventional and learning-based methods", under EPFL's MMPG lab. It performs a comparison between the conventional and learning-based methods for light field denoising. Two sources of noise are considered: Realistic Phone Camera Noise and Additive White Gaussian Noise. The results obtained are shown in the report included in this repository.

Data Set-up Instructions

First, clone the repository. Then, add your own light field scenes to the ./data folder. Two example scenes are already provided The clean images should be in the ./data/clean folder, subvided into folders for each scene. For example, with two scenes named scene1 and scene2 with a grid of dimensions (x1,y1) and (x2,y2) respectively, the folder structure should be as follows:

./data
    ./clean
        ./scene1
            scene1_00_00.png
            scene1_00_01.png
            ...
            scene1_xx_yy.png
        ./scene2
            scene2_00_00.png
            scene2_00_11.png
            ...
            scene2_xx_yy.png

To generate the noisy images, do the following steps:

cd data
python noise_generator.py

By default, the noise generator will generate noisy images with the Realistic Phone Camera Noise model. For other kinds of noise (and other parameters), run the noise_generator.py script with the appropriate arguments. There are three noise categories, depending on the category selected, the images will be placed in the folder ./data/noisy/<noise_category>/<scene_name>.

Methods Set-up Instructions

Wavelet Denoising

No set-up is required for this method. The implemented method is simply a call to SciPy's denoise_wavelet function.

BM3D / LFBM5D

This entire method was adapted from the original LFBM5D implementation GitHub Repo. Go to the method folder via cd ./denoising_methods/classical/LFBM5D. Once in that folder, follow the instructions here to generate the necessary executables.

DnCNN

No set-up is required for this method. The implemented method was taken from the original DnCNN implementation.

LFDnPatch

The folder structure and the training routines are very strongly based on the original original DnCNN implementation, with some modifications to adapt it to the light field domain.

The method relies on an auxiliary repo created for light field patch matching. It is set-up as a submodule in this repository. To set up the submodule, run the following commands:

cd denoising_methods/LFDnPatch
`git clone https://github.com/mmspg/light-field-patch-match.git LF-PatchMatch
cd LF-PatchMatch
git submodule init
git submodule update

Once the submodule is set-up, it needs to be built. To do so, do the following (already inside the LF-PatchMatch folder):

mkdir build
cd build
cmake ..
make

This will create the executable PatchMatch that will be called by the LFDnPatch method.

Usage Instructions

To run the denoising methods on the noisy images, run the following command:

python main.py --scene_name <scene_name> --denoising_method <denoising_method>

There are other possible parameters, but these are the only required ones. To see all the possible parameters, run python main.py --help.

Replicate Results

To replicate the results shown in the report, 3 bash scripts are provided. They are located in the ./scripts folder. However, they are not ready to be run as-is, as the scenes called in the scripts are not all provided. Place your own scenes in the ./data/clean folder, and adjust the scripts accordingly. Use the ./data/noise_generator.py file to generate the noisy images with different noise strengths and types, and the ./data/multi_scale_generator.py to create the downscaled versions of each scene.

Conditions of Use

Permission is hereby granted, without written agreement and without license or royalty fees, to use, copy, modify, and distribute the data provided and its documentation for research purpose only. The data provided may not be commercially distributed. In no event shall the Ecole Polytechnique Fédérale de Lausanne (EPFL) be liable to any party for direct, indirect, special, incidental, or consequential damages arising out of the use of the data and its documentation. The Ecole Polytechnique Fédérale de Lausanne (EPFL) specifically disclaims any warranties. The data provided hereunder is on an “as is” basis and the Ecole Polytechnique Fédérale de Lausanne (EPFL) has no obligation to provide maintenance, support, updates, enhancements, or modifications.

If you wish to use the provided script in your research, we kindly ask you to cite [1].

References

[1] TODO

mmspg/Light-Field-Denoising