[cctalk] Re: OCR and line printers

I too have used Fine Reader (paid for, but the price is not all that high) to OCR various listings. It's way better than the Acrobat OCR, at least the one I tried ages ago on the Ethernet spec (back before Bitsavers, or at least before I found it). The learning feature is very good, and important when dealing with low quality input. You can also tell it only to recognize what you taught it, i.e., not try to match any builtin patterns. If you're doing line printer listings with 64 character sets, that's helpful, otherwise it may mistake something blurry for a pound-sterling sign. In some cases OCR just can't hack what it is given and the only option is to type it all in again. I've done that with some listings that were ugly enough I sometimes had to zoom in to have one letter take up most of the screen, just to figure out what it was. In that particular case, it also used a lot of overprinting to deal with mixed case text: upper case letters represented lower case, upper case with dot overprint for actual upper case. You'd thing that OCR training could handle that, but the difference is subtle enough it doesn't really work. Supposedly current versions of the open source program Tesseract are pretty good, but I haven't tried it. Looking for how to train it got me all confused, it didn't seem to be something that was at all convenient, not like the interactive training feature of Fine Reader. OCR likely will not handle fixed layout well (not unless you can treat it as tables). If that's important, some Python or Emacs post-processing can clean up a lot. Similarly if there are common recognition errors you can spot by pattern matching. Scanning the CDC 6600 wire lists goes well this way, because the data have a very consistent pattern. For example, the OCR might mix up zero and oh, but an edit pass can fix those 100%. paul