I worked on this project for a class at Missouri S&T called Machine Vision. For the project I recreated some of the work done in [1] comparing image compression techniques to illustrate the benefits of compressive sampling. The technique demonstrated in [1] showed that random noiselets could be used for compression using l1 minimization recovery, and that this technique could actually yield better visual results than the typical discrete cosine transform alone. Oddly enough I showed that the benefits of using this technique did not hold when tested on images other than the demonstration used in the article. I was certainly surprised by this result!
[1] Romberg, J. (2008). Imaging via Compressive Sampling. IEEE Signal Processing Magazine, 25(2), 14{20. doi:10.1109/MSP.2007.914729