imageProcessing | Survey of Image Denoising Techniques

Howdy! If you're reading this it means that you've stumbled across a bit of code that I've written as a class project.

The code in this repo is designed to perform a variety of image denoising techniques in CUDA. This code is intended to add noise to clean base images, and then remove it using some of the included methods--however, it can be used on images without adding noise by setting the <noise_rate> paramater to 0.

The code should be built using cmake, but a manual makefile is also included for those on Vocareum.

The primary built target is called imageProcessing, and it contains a few methods that

• add additive white gaussian noise
• add salt and pepper noise
• do a median filter
• do svd on an image
• do a fourier low pass filter
• do variational diffusion
• do total variation minimization
• and more!

Python scripts are included to extract(pixelExtractor) images into a .pxa (pixel array) format, and compress(pixelCompressor) .pxa files back into .png or other image formats.

Good luck. Hopefully this code is helpful useful to you.

Building the code

cmake version 3.21 or later is required to build the code for this project. If you are in the Vocareum enviroment, a MANUALMakefile has been provided, and should be used. To install the required Python 3 packages, and compile the main code with cmake run the following:

cd /.../.../imageProcessing/
pip install -r requirements.txt # may need to be pip3 
mkdir build
cd build
cmake ..
make -j

Ta da! Hopefully everything built correctly.

If you are on Vocareum, running a simple:

cd /.../.../imageProcessing/
pip install -r requirements.txt # may need to be pip3 (NOTE: vocareum should already have the python packages installed, should be okay if this fails)
mv MANUALMakefile Makefile
make -j

Should do the trick.

Demo usage and examples

I recommend using only smaller images (up to 400x400), as it can be tricky to see the effects with lots and lots of pixels.

If you do not have a testing image--images/camera.png, images/butterfly.png, images/moon.png, or images/barn.png--can be used by copying them to your working/build directory.

If you need to convert an image to .pxa format, you can run python3 pixelExtractor.py inputfilename.png to get a inputfilename.pxa file. Note: Most file formats are supported via the Python Image Library (PIL), so you can use .jpeg, .png, etc.

To just add additive white gaussian noise w/ std. deviation 20 ./imageProcessing inputfilename.pxa outputfilename.out 1 0 20

To just add salt and pepper noise w/ rate 30 ./imageProcessing inputfilename.pxa outputfilename.out 2 0 30

To perform total variational reduction on an image with additive white gaussian noise w/ std. deviation 20 using lambda .1 ./imageProcessing inputfilename.pxa outputfilename.out 1 6 20 .1

To perform a median filter on an image with salt and pepper noise w/ rate 30 using lambda 1 ./imageProcessing inputfilename.pxa outputfilename.out 2 4 30 1

To view the processed images you can use python3 pixelCompressor.py outputfilename.out outputfilename.png to get a finished image.