The following is a short analysis of the recognition scores of the new system. The 10-fold cross-validation shows that about 85% of the values is recognized. This score differs between 80% and 90% for each time I run the cross-validation so this means there are not enough samples. Analyzing the folds themselves, it shows that values can’t be recognized because the digits that are incorrectly estimated were not in the training set.

If the 100% score is feasible, remains to be seen (need more samples), but my hopes are high :)