spamsickle wrote:I've run into situations where I had difficulty reading tiny subscripts here and there shooting one page with 8 MP, but so far I've been able to puzzle out what Scan Tailor drops out when doing its binarization without needing to consult the original JPEGs.
I contend that a good scanning process should make even the smallest type fully legible and OCR of said text nearly perfect.
The following assumes that binarization means production of b&w images, vs. leaving images as greyscale or color.
Thus, one question -- must Scan Tailor binarize to function properly?
-- and one, long FYI, if it makes any sense: my experience using Acrobat OCR (perhaps dedicated OCR would be different) is that binarization of originally greyscale or color images yields poorer OCR.
This opinion is based on just a handful of tests I ran on book material I am ongoingly scanning with a Plustek Opticbook flatbed scanner (until I settle on and build a diy machine based on the profusion of material available on this site). Since it was just a handful of tests, in the end this observation about OCR of b&w images may not be a scientific fact, but . . .
I used to scan all pure text pages in b&w, and only book pages with images in greyscale or color, since the former took up just 1.5Mb per image, the latter a great deal more at 8Mb. This system was necessary a few years ago when hard disk space was a factor in things, but it also made intuitive sense -- b&w images just look more OCR-able. The follow-on OCR was most conveniently done in Acrobat although it was not that great, which I accepted as worth the convenience of having OCR embedded in page images. It never occurred to me that the dirtier greyscale images could possibly be easier to OCR.
Anyway, a year or so ago I happened to both accidentally scan some of a book's pure text pages in greyscale and output the resulting Acrobat OCR to a Word file (usually I am content to leave the OCR content embedded in the Acrobat file). Some parts of the export had obviously fewer errors and on investigating the source images I discovered that the greyscale-scanned pages were substantially less prone to errors than the b&w-scanned pages. Two more contrasting scanning efforts confirmed this observation.
A look at the different kinds of images in extreme closeup (I think I blew them up to 400%), both in an imaging program and in Acrobat (the difference is almost undetectable, which is amazing given the reduction in aggregate file size achieved by the Acrobat combining utility) shows a possible reason why -- binarization results in little bits being cut out of or added to the letters, bits of the letter or the background respectively that the binarization software cannot tell is not part of the letter, bits that the human eye adds in or removes -- or never deals with to begin with -- when looking at greyscale versions of those same letters embedded in their dirty greyscale "white" background.
Now, it does not really make sense that binarization elements of the scanner software (or the binarization elements of Scan Tailor) would have difficulty figuring what is and is not part of a letter from a given source when at the same time, working on the same or even slightly degraded material, simple Acrobat OCR software has less difficulty doing so. Both processes must distinguish letter from non-pure-white background. I would assume that the underlying algorithms for determining what is and is not part of a letter against a noisy background are long established and more or less common to all this software -- the great OCR programs like Omnipage haven't significantly increased their recognition rate in many years and only Google seems to be able maintain a software technology edge for the long term. But as far as I can tell, it is exactly the case that binarization software does not do as good a job as even run-of-the-mill OCR software (the OCR engine in Acrobat can't be the best) in deciding what is a letter.
I feel like my observations must be in error since the logic of the binarizing / OCRing software in the two scenarios says there shouldn't be much significant difference. But for now I am going with the observations of divergent OCR accuracy and I both scan all books in greyscale (for DIY scanning I assume color or greyscale outputs are default) and leave the final images in the finished ebook product in that state both for image readability and superior embedded OCR. When I move to a DIY scanner I will want to do the same.
And just to reiterate the starting question -- must Scan Tailor binarize to function properly? It does not seem to have any setting to allow it to function otherwise.
