I don't know about iTextSharp, but you have to rewrite a PDF file if anything is changed, as it contains an xref table (index) with the exact file position of each object. This means if even one byte is added or removed, the PDF becomes corrupted.
Your best bet for recompressing the images is JBIG2 if they are B&W, or JPEG2000 otherwise, for which Jasper library will happily encode JPEG2000 codestreams for placement into PDF files at whatever quality you so desire.
If it were me I'd do it all from code without the PDF libraries. Just find all images (anything between stream and endstream after an occurance of JPXDecode (JPEG2000), JBIG2Decode (JBIG2) or DCTDecode (JPEG)) pull that out, reencode it with Jasper, then stick it back in again and update the xref table.
To update the xref table, find the positions of each object (starting 00001 0 obj) and just update the new positions in the xref table. It's not too much work, less than it sounds. You might be able to get all the offsets with a single regular expression (I'm not a C# programmer, but in PHP it would be that simple.)
Then finally update the value of the startxref tag in the trailer with the offset of the beginning of the xref table (where it says xref in the file).
Otherwise you'll end up decoding the entire PDF and rewriting it all, which will be slow, and you might lose something along the way.
