[Bug]: Every PDF I OCR has the text misaligned with the image about ocrmypdf

[jbarlow83]

@Mark-Joy Thanks, it looks like it's actually just changing -1 to

-text.text_transform(Matrix(1, 0, 0, 1, box.llx, 0))
+text.text_transform(Matrix(1, 0, 0, -1, box.llx, 0))

Translation: the invisible font is upside down, and some PDF viewers freak out. 🙃

But I need to test it everywhere first, since the first one slipped through.

v16.0.2 temporarily makes sandwich the main renderer again.

[Mark-Joy]

A quick fix shall be:

Change:

OCRmyPDF/src/ocrmypdf/hocrtransform/_hocr.py

Line 290 in 0b6fb62

bottom_left_corner = line_box.llx, line_box.ury

to bottom_left_corner = line_box.llx, line_box.lly

And:

OCRmyPDF/src/ocrmypdf/hocrtransform/_hocr.py

Line 312 in 0b6fb62

fontsize = line_box_height + intercept

to fontsize = line_box_height - intercept

[jbarlow83]

16.0.2 is a temporary fix - I'll close this issue when there's a full solution and the new renderer can be reinstated as default.

The sandwich renderer (default in 16.0.2 and <16) has a number of issues like wordsegmentationproblemsinsomecases and registration (aligning selected text to actual). The new renderer, which is mostly a reimplementation of sandwich to fix sandwich's issues, doesn't work universally yet, but does have significant improvements.

[Rosti2022]

Thanks for 16.0.2 - I can confirm that I am witnessing the same issue with my PDFs, as described in #1214. The Persian text issues persist in 16.0.2 unfortunately.

[advert665]

Thanks all -- v16.0.2 fixes the issue for me, very pleased!

[Pugio]

Thanks for the fix. Can confirm that this bug caused several hours of head scratching today before I saw this issue. Really appreciate all the great work!

EDIT: I was also seeing a huge number of \ufeff characters in the output. Moving from 16.0.0 -> 16.0.2 also fixed this.

[jbarlow83]

v16.0.4 should contain significant fixes when running in --output-type hocr mode. It has not made been made the default.