Difference in word coordinate information

[yavuzKomecoglu]

Hello, in some Turkish newspaper PDFs, the word coordinates are below a certain pixel. For example, while the headlines in the news in 2_eylul_15_11_2018.pdf / page-3 and 2_eylul_30_11_2018.pdf / page-7 start from the bottom, the coordinates are obtained correctly in the news in gaziantep_dogus_10_08_2021.pdf / page-4.
What exactly is the difference between them? Why does this occur? What's the difference with these pdf's that such coordinate information returns differently?
Thanks.

First, the words are extracted.

data_words = crop.dedupe_chars().extract_words(x_tolerance=1, y_tolerance=1, extra_attrs=["fontname", "size"])           

Then the title region is determined.

h_left_x0_all = min([h["x0"] for h in title])
h_left_top_all = min([h["top"] for h in title])
h_right_x1_all = max([h["x1"] for h in title])
h_right_bottom_all = max([h["bottom"] for h in title])

title_area = (h_left_x0_all, h_left_top_all), (h_right_x1_all, h_right_bottom_all)

Note: Where title_area is absolute position, relative position is used when drawing title area with opencv.

Test Newspapers

• 2_eylul_15_11_2018.pdf
• 2_eylul_30_11_2018.pdf
• gaziantep_dogus_10_08_2021.pdf

[jsvine]

Hi @yavuzKomecoglu, and thanks for your interest in this library. Your examples and questions are very interesting ones. Unfortunately, I don't have a very conclusive answers. My best guess at what's happening is either:

• The PDF's font definitions are unusual or incorrect. If that's the case, I'm not sure there is much you could do to fix that.

• pdfminer.six's coordinate-calculation logic is flawed for those fonts. (pdfminer.six is the library we use to extract all character coordinates, from which we calculate the word coordinates.) If that's the case, it may be worth filing a discussion or issue on that repository.

My apologies that I can't be more helpful, but feel free to continue the discussion. Perhaps someone else here has deeper insights into what's happening.

[yavuzKomecoglu]

Thanks for your feedback.
in this case it only occurs in some header objects in this pdf specific.
As a solution suggestion, I tried to determine the font metrics values of PIL and the blank values from above and below according to different font sizes. This is partially a solution.