In 2023, I enrolled in a graduate-level course on machine learning and statistics taught by Professor Qi Guo at Purdue University. The course required us to undertake a final project that involved re-implementing a seminal machine learning publication. I chose to focus on the paper titled "Plug and Play ADMM for Image Restoration," authored by Dr. Stanley Chan.
My efforts in this project resulted in a grade of 86.0/84.0 for the final project and an overall A in the class. I have compiled my work on this final project, along with the paper, in this GitHub repository.
Since this is a re-implmentation of a very well documented project, there is an existing codebase for the original project here. This github does three things differently than this original codebase:
- That codebase is in MatLab while this one is in Python3
- The parameters are a bit different - not better or worse but it's a re-implmementation in a different language so the values are different
- This project expands upon the old one and tests results for different scenarios than the original which only tested on a few
- The PNSR is lower in general for this project because the noise is higher than in the MatLab demo
To see the ADMM algorithm in action, go to the demonstrations folder. All scripts can be run with python and will display the before and after denoised image. You should run the files from within the demonstrations folder.
In the final project document, there are graphs displaying how the Peak to Signal Noise Ratio (PNSR) changes given different parameters. This is calculated with the scripts in the analysis folder. These scripts take a long time, sometimes several days, to run. I highly reccomend using other computing recourses. Also, the files can load python pickles from the pickles folder if you don't want to generate the data from scratch. The generated plots are also in the analysis folder.