Classification and its difficulties

We have started building our training set.  However, we are running into some difficulties on the way.  One of the main problems is that the connected component analysis did not work as well as we hoped it would.  Some of the note symbols would be stuck together, while others would be much too segmented.  Nevertheless, we still tried labeling some of these symbols and running a k-NN algorithm on the rest of the unlabeled symbols and seeing what kind of results we would get.  The results were not too bad; it was not complete noise, but it was definitely not as accurate as it could be.  To improve the results, one thing we can try is increasing the amount of manually labeled symbols included in the training data.