dfridovi / compressed_sensing

Image compression using compressed sensing.

compressed_sensing

Image compression using compressed sensing.

Summary

This repository is under development as part of a class project for UC Berkeley's EE227BT Convex Optimization course. The authors are David Fridovich-Keil and Grace Kuo, both graduate students in the EECS department at UC Berkeley.

Organization

The files in this repository are organized as follows.

The compressed_sensing/presentation directory contains a copy of our slide deck, and also several images used in the slides.

The compressed_sensing/writeup directory contains a copy of our final report.

The compressed_sensing/data directory contains three example images. Virtually all of our examples in the slides and the report use the lenna.png image.

The compressed_sensing/reconstructions directory contains two sub-directories, matlab figures and python figures, which (not suprisingly) contain compression and reconstruction results created by test scripts written in MATLAB and Python, respectively.

The compressed_sensing/src directory also contains two sub-directories. The matlab sub-directory contains our most up-to-date code base; these are the functions and scripts we use to generate all the figures in our presentation and report. The python sub-directory contains an earlier version of the code base.

compressed_sensing's Issues

Fourier

Replace all Fourier functionality with something that only uses real numbers, like DCT. This will allow us to apply the same L1 regularization and other CS techniques in the transform domain, which might improve compression.

Add reversed Huber norm relaxation

As in problem set 3, replace the L1 regularization with reversed Huber norm. Try this in the image domain, and also in the transform domain.