As a comparison to the much simpler Hu moments, Chen's method improved letter recognition by a significant 25 percentage points (65% vs 90%). Shape recognition was unchanged.
Our final report is here: http://cseweb.ucsd.edu/classes/wi10/cse190-a/reports/wgrant_lkander.pdf
We may post the matlab code for Chen's method, which involved radon transforms, dual-tree complex wavelet transforms, and fourier transforms, as well as k-means clustering for image segmentation.
Any questions, just send a message.
Also, we got an A. :)