Made this roughly a year ago (see <a href="https://news.ycombinator.com/item?id=9652913" rel="nofollow">https://news.ycombinator.com/item?id=9652913</a>).<p>Originally I let the neural net output a charcode array (either 4 or 8 bits, depending on if the output was a number or a character) which could then be converted into the corresponding character. In short, this improved performance and decreased accuracy. I ended up switching back to standard mapping (index of maximum probability is the best guess), which requires an output layer the same size as the amount of characters one wants to recognize.<p>Posted it here again because the demand for this kind of tool still seems to be quite high and there is no FOSS alternative that I know of.<p>If you have questions or suggestions, let me know!