Text Bounding Box Below Actual Text Position

[renraeldab]

I guess this can happen when there are some uncommon fonts. Is there any way to detect or fix this?

[JorjMcKie]

No, this cannot be detected: we are confined to believe what a font is telling us ...
It may be a bug though - the picture does look a bit weird. It might be worthwhile to take a look at the page.
You could marks the rectangles based on extraction data after you have executed pymupdf.TOOLS.set_small_glyph_heights(True). This option causes PyMuPDF's character bbox computations to take a different path.
If this phenomenon then goes away, we do have a bug (in MuPDF).

[JorjMcKie]

Ok, thanks.
Please let us have the PDF page. I would like to have a look at it.

[renraeldab]

Link to the pdf: https://iiis.tsinghua.edu.cn/index.php?f=down&v=d&s=T4HPwJkpEXiUvs1bOwgbD6x2Md7WKMrrZWjYQHDPNOvrZhh6RmhU5S007nqjFuXriXO72831ol7QbJlWKuzehqwSIDcdN%2BwH

[JorjMcKie]

Thanks for the file.
I looked at the used fonts and found one that is erroneous in that it contains crazy values for the so-called ascender and descender values. Please look at the definition of these properties here.
Both are crucial for computing text positions. Ascender must be positive and descender must be negative. This is what we find:

for itm in page.get_fonts():
    buff=doc.extract_font(itm[0])[-1]
    font=pymupdf.Font(fontbuffer=buff)
    print(f"xref {itm[0]}, {font.name=}, {font.ascender=}, {font.descender=}")

    
xref 144, font.name='AdvOT735b620e Regular', font.ascender=1.0, font.descender=-0.29100000858306885
xref 145, font.name='AdvOT144a03c1 Regular', font.ascender=-32.768001556396484, font.descender=-0.25999999046325684
xref 146, font.name='AdvOT5e4d79fc Regular', font.ascender=0.8539999723434448, font.descender=-0.25
xref 155, font.name='AdvOT1ef757c0 Regular', font.ascender=0.8539999723434448, font.descender=-0.25
xref 165, font.name='AdvTT16f3b945.B Regular', font.ascender=0.9580000042915344, font.descender=-0.25
xref 166, font.name='AdvOT1415ea69.B Regular', font.ascender=1.0, font.descender=-0.25999999046325684
xref 167, font.name='AdvOTf32ed098.B Regular', font.ascender=1.0099999904632568, font.descender=-0.25999999046325684

Font at xref 145 shows an ascender value of ridiculous -32.768. Unacceptable for position computations: some reasonable values must be used instead.
Unfortunately, this corrective algorithm does not work satisfactorily here ...

[renraeldab]

The ascenders and descenders are all normal in get_text("dict") results, different from the fonts info. Is this after correction?

[JorjMcKie]

Yes - but the subsequent code does not correctly used the corrected ascender / descender values.
This is in contrast to page.get_texttrace() - a method acting independently of a TextPage. The per-character boundary boxes here are sufficiently corrected.

[JorjMcKie]

I think that MuPDF makes an error here. After confirmation, I will submit a bug into their system.