Many of the published techniques we uncovered in our library research for this project were found to be "overkill". Matlab routines exist and proved useful for several of the required procedures, such as thinning and implementing the neural network. An abundance of useful information was also found on the web relating to segmentation, slant correction, pattern recognition, etc.
Thanks to the Matlab toolboxes, exercises such as thinning proved to be faster than expected to implement. Determining unique parameters for distinguishing characters was fairly difficult as there was a lot of overlap between the various digits.
Our neural network was trained using "noisy", handwritten characters only. This facilitates recognition of handwritten digits, but reduces the system's ability to handle typeface. Training the system is a resource-intensive exercise. Had the system been trained with both noisy and clean digits, as well as a larger library of noisy digits, we would expect more accurate results.
Overall this was an interesting project with potential for real applications. In hindsight, due to the time and memory constraints, we are glad we chose to tackle only ten numerical digits, rather than 26 alpha characters. We would expect to have had better success in our classification results if we had analyzed typeset data rather than variable handwritten data.
The task of identifying characters based on features was a complete success. Really and truly, A+ material. *