Question: Are JBIG2 file compression and document OCR completely separate problems? Should these processes be done together or separately? Does JBIG2 conversion prior to OCR lower the recognition rates?
Answer: JBIG2 and OCR are related problems. A basic element of JBIG2 compression is bottom-up font learning. This font learning is used for compression but can easily be used to cross-check font mappings returned by the OCR engine. So an effective JBIG2 compression algorithm can be used to improve on OCR recognition rates, see http://www.cvisiontech.com/pdf_compressor_31.html.
For example, if we use global models in JBIG2 which is an effective compression tool, it can also be used to propagate correct OCR mappings throughout the document.
In general, these processes should not be constructed in a linked way, and each process, JBIG2 & OCR, needs to be able to run without the other. One important reason for process separability is speed: JBIG2 tends to runs at 3-5 pages/sec. while OCR can take 5 secs a page. Another reason is that to achieve good OCR rates the right language dictionary needs to be used. JBIG2 compression is language independent and should not rely on any language dependencies.
So the JBIG2 and OCR problems are certainly very-much interrelated. Having said that, there are many reasons (including speed) to solve them separately and then combine results. Certainly, there should be an integration phase where a higher level module is aware of both the JBIG2 and OCR results and is able to combine these results to acheive improved OCR (and maybe also JBIG2) results.
There are problems inherent in propagating OCR results across a document. If there are any OCR errors these can also propagate across the document with negative consequences. Obviously, it is important in any such fusion of OCR and JBIG2 results to make sure this kind of error propagation does not occur.
Reliable JBIG2 compression, done with precision, should not result in any degradation of the document. As such, JBIG2 conversion prior to OCR should not lower recognition rates. There are, however, JBIG2 compression implementations that are clearly lossy and degrading in nature. If one of these degrading JBIG2 methods is run prior to OCR, a drop in recognition rates can usually be expected.