• Welcome to the new Internet Infidels Discussion Board, formerly Talk Freethought.

Optical Character Recognition (OCR) Software

lpetrich

lpetrich

Contributor
Joined
Jul 27, 2000
Messages
26,343
Location
Eugene, OR
Gender
Male
Basic Beliefs
Atheist
Has anyone used OCR very much?

I've used OCR over the years, like to transcribe a big collection of numbers from a journal article. I mainly used GOCR/JOCR, Tesseract, and ReadIRIS. The first two are open-source, and the third is commercial -- it came with a scanner that I once owned.

I recently got OwlOCR in the Mac App Store. It's good. I've used it on some of the image-of-text files that I find in Twitter. I have to edit the resulting text afterwards, because it doesn't try to guess what's part of a paragraph.

OwlOCR has three sources of image input:
  • Image file - from the filesystem
  • Screenshot - drag-and-drop selection
  • Scanner and camera
It doesn't have image copy-and-paste, however.

It also has "Language Correction", but that is its only interpretation option.
 
A potential problem with OCR is what range of possible interpretations that one can use. If one uses a smaller set of possible interpretations, then each one can tolerate a poorer fit before it becomes hard to distinguish from another one. But a smaller set may not correctly represent the source text.


This can be illustrated with Bayes's inverse-probability theorem. For data D and hypothesis H, one has probability

P(H if D) = P(D if H) * P(H) / P(D)

P(D) is found by making P(H if D) add up to 1 for all possible values of H.

So it's approximately (hypothesis given data) = (data given hypothesis) * (hypothesis by itself)

The big problem is what values of P(H) to use, and that's what makes Bayesian statistics so controversial.

If one allows only one possible hypothesis, H1 by making only P(H1) nonzero out of all the P(H)'s, then only P(H1 if D) will be nonzero out of of all the P(H if D) values. So one has "confirmed" H1. Likewise for more than one hypothesis, and weighting toward one hypothesis rather than another.


Applying this to OCR, if one uses the hypothesis of only horizontal-line ASCII characters, one will only get such characters, even when others are present in the image. But having more character sets and more orientations means being able to cover those cases, though at the risk of misidentifications. So some good OCR software ought to have the option of which character sets to use, and which disambiguation strategies, like for letter O vs. numeral 0, and for letter I vs. letter l vs. numeral 1 vs. punctuation character !.
 
Near instant, high quality OCR for Mac with OwlOCR - OwlOCR's home page.

I've found several other OCR apps in the Mac App store, and one can also find OCR apps for Windows, Linux, iOS, and Android.

I've also found several online OCR sites. It's those sites that do the work, so one doesn't have to download or install anything.

There are some OCR sites that do other writing systems: Top 10+ Free Online Russian OCR software of 2023

Some sites do several:
Roman, Greek, Cyrillic, Hebrew, Arabic, Syriac, Ge'ez, Armenian, Georgian, Mongolian, Chinese, Japanese, Korean, South Asian (Tibetan, Devanagari, Bengali, Assamese, Gujarati, Odia, Kannada, Malayalam, Tamil, Telugu, Sinhala), Burmese, Thai, Lao, Khmer, Javanese, Sundanese, Cherokee, Inuktitut
 
You must log in or register to reply here.
Back
Top Bottom