Tesseract Teaser

Reminds me of an issue I had with the Python Pillow library had generating checkerboard patterns. I spend ages getting smaller and smaller versions of the problem until I could reproduce it in 5 lines of very clear code[1].

I was convinced I was doing something wrong with pixel access, or many the mode of the image.

Turned out the solution was "upgrade Pillow". Ironically - and annoyingly - both my question and answer were downvoted.

[1] https://stackoverflow.com/q/63256952/280795

There's a bunch of ways this might be broken, anything from user error, alpha channel, bug in get.images, etc. There's a suggestion that the forums are the right place to post this sort of thing at: https://github.com/tesseract-ocr/tessdoc/blob/master/Improve...

The fact that the title of this article contains "teaser" suggests there might be a second part that explains the cause... but based on the symptoms so far, bug is almost certainly in the preprocessing stage.

I happen to have an extremely old version of Tesseract that I used in a project long ago, so out of curiosity decided to try it on the article's image. It's old enough that it didn't support anything but 8bpp or less BMP input, so I had to convert the PNG to a 256-color BMP, which resulted in the following output when I fed it to Tesseract:

    TUE 03/@4/2018
    BBC TWO
    1W 8:30
    Oh30m (mt]

Pretty miserable result, but no mysterious extra line either.

I wouldn't discard a more traditional error like memory corruption or some buffer overflow

When is this from? I just tried the supplied "Original" image in Tesseract 4.1.1 (installed from MacPorts) and it worked flawlessly:

    TUE 03/04/2018
    BBC TWO

    18:30

    0h30m (DR)

Edit: Apparently this is from just last week (2021-03-06) https://www.solipsys.co.uk/new/RecentChanges.html

I'm so confused by Tesseract. With the correct language specific model the results are mostly flawless but then sometimes it goes like "...,,..,,[]////gtghth".

Almost sounds like an Underhanded C Contest story.