Thanks Ger, this has been incredibly helpful! Reducing the image size for OCR has dramatically increased the accuracy and reliability of my output.
On Wed, Feb 22, 2023 at 11:31 AM Ger Hobbelt <g...@hobbelt.com> wrote: > Re the line pixel height research I mentioned I recalled: it's here: > https://willus.com/blog.shtml?tesseract_accuracy and here: > https://groups.google.com/g/tesseract-ocr/c/Wdh_JJwnw94 > > I had forgotten I got it from this mailinglist! > > On Sat, Feb 18, 2023, 18:39 Ger Hobbelt <g...@hobbelt.com> wrote: > >> Hi, >> >> Had a very quick look but got sidetracked into something else, so I >> didn't write the tesseract test script I wanted, so TILAAEFTR. Here goes: >> >> your '4' output image is rather large for tesseract to treat it as a >> 'single line'. >> >> tess is known to deliver different accuracies for (*wildly*) different >> line sizes -- I seem to recall some research and graphs from 2019 where >> accuracy went down for both too small (8-10px) and *way too high* (200+px), >> producing a bit of /skewed/ bathtub curve for the OCR error rate, so the >> idea here is to rescale your extracted number images to a suitable size, >> before feeding it ot the OCR engine. >> >> Test this remark/idea with a script: >> >> ``` >> let img = 'out.png' // the '4', f.e. >> for (let h = 8; h < 500; h = ceil( h * 1.1 /* = +10% */ )) { >> /* use imagemagick for scaling, f.e.? */ >> rescale(img, height: h, unit: 'px') -> img2 >> tesseract(img2) -> txt >> } >> ``` >> >> (pseudocode above; write in your favorite scripting language: bash, js, >> python, whatever) >> >> collect the `txt` OCR results; rank them and see where your 'optimum >> height' lands you. Then use that for your application. >> >> >> >> Afterthought / Side thought: >> >> I see you are grabbing a computer display screen and applying OCR to it. >> A few thoughts pop up immediately given the source type: >> >> I see a rather organized screen, no noisy/chaotic background you get with >> burned-in subtitles, for example. Food for thought. >> >> - doesn't it suffice to take the number (*digit*) images and compare them >> against a (created) master set, using a image similarity metric? As it's >> the machine rendering those numbers, they should be pretty consistent, save >> for some anti-aliasing or non-pixel-accurate positioning in the renderer >> resulting in (slightly) different pixel values / images for each digit. >> (Feels like tesseract is an elephant gun for this. But then I probably >> missed several cues and be utterly wrong...) >> >> - of that same vein, taking it one further: since it's output from a >> computer machine, can't we hook into the software which produces these >> images and get the raw digital numeric / scoreboard data from the software >> straight away? Iff we can, we don't have the significant overhead and data >> accuracy challenges that come with reversing anything using OCR: it's never >> a 100% accuracy this way. (software protections and other obstructions >> related to data commerce and ~ politics can keep us at a distance, where >> screengrabbing+OCR becomes an optimum viable solution if we want to get >> access to the data, but I would love to get away with less for the same (or >> better) result. :-S ) >> >> - is it me or am I seeing more of this machine -> >> screengrab/scan/photograph, digitally or *analog* (phone snaps of other >> phones' screens) -> machine OCR data transport queries lately ('22 / '23)? >> Have I missed something? >> >> This looks like trade/score screens and at least the traders would have >> *some* incentive to provide an API for this. (When you find the related >> paywall insurmountable, grab+OCR is the way to go, alas, but it will always >> be somewhat finicky.) >> >> >> >> Met vriendelijke groeten / Best regards, >> >> Ger Hobbelt >> >> -------------------------------------------------- >> web: http://www.hobbelt.com/ >> http://www.hebbut.net/ >> mail: g...@hobbelt.com >> mobile: +31-6-11 120 978 >> -------------------------------------------------- >> >> >> On Fri, Feb 17, 2023 at 8:08 AM Alex Porter <trumpetdu...@gmail.com> >> wrote: >> >>> am currently building a pythont tool to read the screenshots of a >>> in-game scoreboard. The scoreboard looks like this:[image: ss_1.png] >>> >>> I am using open cv to analyse the scoreboard and can reliably slice the >>> image into rows and extra each value from the scoreboard giving an image, >>> after processing, like this:[image: crop3.png] >>> >>> I am still having issues with tesseract accurately identifying the >>> numbers. Sometimes it is inaccurate (identifying the wrong number) or not >>> giving any output at all. I have only whitelisted 0-9 when reading the >>> numbers. Any help on pre-processing the image to increase accuracy or any >>> other ideas would be much appreciated! >>> >>> I have also attatched the python code. It's quite messy in it's current >>> form so please forgive that if you decide to look! >>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "tesseract-ocr" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to tesseract-ocr+unsubscr...@googlegroups.com. >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/tesseract-ocr/19961d38-af02-4253-801d-4de53493cf54n%40googlegroups.com >>> <https://groups.google.com/d/msgid/tesseract-ocr/19961d38-af02-4253-801d-4de53493cf54n%40googlegroups.com?utm_medium=email&utm_source=footer> >>> . >>> >> -- > You received this message because you are subscribed to a topic in the > Google Groups "tesseract-ocr" group. > To unsubscribe from this topic, visit > https://groups.google.com/d/topic/tesseract-ocr/jWdpUF7mTxE/unsubscribe. > To unsubscribe from this group and all its topics, send an email to > tesseract-ocr+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/tesseract-ocr/CAFP60fpxFVty%2Ba66Ndhb258rggN4u4OY%3DC62asW9_j3%2BoNzFAw%40mail.gmail.com > <https://groups.google.com/d/msgid/tesseract-ocr/CAFP60fpxFVty%2Ba66Ndhb258rggN4u4OY%3DC62asW9_j3%2BoNzFAw%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-ocr+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/CABVhM9BqV2kuUnP0E3XR2WtNpkiO1BBScerjTZPhcZYg7zZWOA%40mail.gmail.com.
