Thank you Bruno Martin for your reply. Basically we're working on DIBCO 2011, 12, 13 databases. Please see the link for 2011 database (http://utopia.duth.gr/~ipratika/DIBCO2011/benchmark/dataset/).
You seem to have difficult documents in which we also see the backside through the paper. Some of these cases have been tackled with hyper-spectral methods combining infrared and visible information but they are not applicable in case you only have a high resolution picture of the document and not the original to work on.
The last two links seems to offer good algorithms to work with your type of documents...