I am getting nearly 99% accuracy by training from the top layer of the network. I am training using synthetic data; and the evaluation is done the same type of data.
But, the result is not extending to actually scanned documents. On the scanned documents, I am getting lower accuracy, unfortunately.
So, it looks like training using synthetic data is actually not that productive approach.