Issue with coordinates of text and table when cropbox is smaller than mediabox and rotated

[AllisonHu64]

Description of the bug

Based on my research, Mediabox defines size of the pdf page. Cropbox defines the rect of the page displayed by PDF Viewers. Pixmap displays the area intersected with Cropbox and Clip. And from my understanding and tests, the coordinates of text retrieved by functions such as Page.get_text() is with respect to the Cropbox and unrotated. (Pymupdf should have better documentation regarding the coordinates.) I recently discovered that after using the function Page.find_tables() on a rotated page where cropbox is smaller than mediabox. The Cropbox value changes, resulting subsequent Page.get_text() coordinates to change.

Can someone please take a look? Thank you so much!

How to reproduce the bug

Please download the test pdf attached.
Install the following pymupdf in a clean virual environment
python -m venv venv pip install pymupdf=="1.25.3"

Run the following code
`
import fitz

# Read the pdf file
pdf_document = fitz.open("test.pdf")
page = pdf_document.load_page(0)
print("Before: ", page.cropbox, page.cropbox_position, page.rotation_matrix)
print("Before: ", page.search_for("第七章"))

print("\n")
# Find the tables
tables = page.find_tables()
print("After: ", page.cropbox, page.cropbox_position, page.rotation_matrix)
print("After: ", page.search_for("第七章"))
`

The result will look something like this. The page's Cropbox and its related property changes after running page.find_tables
`
Before: Rect(36.0, 36.0, 559.0, 805.9000244140625) Point(36.0, 36.0) Matrix(0.0, 1.0, -1.0, 0.0, 769.9000244140625, 0.0)
Before: [Rect(194.8800048828125, 38.02104568481445, 237.0097198486328, 52.66114807128906)]

After: Rect(0.0, 0.0, 595.2999877929688, 841.9000244140625) Point(0.0, 0.0) Matrix(0.0, 1.0, -1.0, 0.0, 841.9000244140625, 0.0)
After: [Rect(230.8800048828125, 74.02104187011719, 273.00970458984375, 88.66114807128906)]
`

test.pdf

PyMuPDF version

1.25.3

Operating system

MacOS

Python version

3.12

[JorjMcKie]

This is working as intended!
Page.find_tables() will de-rotate the page via page.remove_rotation() before locating tables.
Therefore all later extractions of page content will / may show different results.

[AllisonHu64]

This is working as intended! Page.find_tables() will de-rotate the page via page.remove_rotation() before locating tables. Therefore all later extractions of page content will / may show different results.

Hi, I am not talking about rotation. It seems likePage.find_tables() changes the coordinates of the Cropbox. I don't think that is supposed to happen. As you can see, In the demo code i provided, the coordinates of the text changes from [194, 38, 237, 52] to [230, 74, 273, 88] which is an offset of [36, 36]. The Page.find_tables() changes the origin of Cropbox from [36, 36] to [0, 0] making it the same size as the Mediabox.

[JorjMcKie]

If there is a rotation != 0, find_tables() will indeed change MediaBox and CropBox. Hopefully correctly computing things ... but the change is intentional.
The result is an identical looking page with rotation 0 that behaves like normal.
I will check your example to see if I am miscalculating things though.

[AllisonHu64]

Thank you so much for double checking. But I am wondering if page.remove_rotation() and changing the Mediabox and Cropbox is necessary.

According to the documentation website (https://pymupdf.readthedocs.io/en/latest/page.html#modifying-pages),

So the returned value of e.g. Page.get_image_bbox() will not change if you do a Page.set_rotation(). The same is true for coordinates returned by Page.get_text(), annotation rectangles, and so on.

Since page.get_text() and page.get_drawings() (I wrote some code and found page.get_drawings() coordinates stayed the same before and after rotation) isn't affected by page.remove_rotation. And page.find_tables() uses page.get_text() and page.get_drawings() to find its coordinates. See Code (https://github.com/pymupdf/PyMuPDF/blob/main/src/table.py#L1867)

Can we remove page.remove_rotation and the code related to Cropbox before find_tables?

[JorjMcKie]

Thought about this a bit:

• The Table Finder requires horizontal text.
• If the page appearance shows horizontal text although the page is rotated, we therefore need page.remove_rotation(). This can be detected by page.rotation != 0 and all (or most) lines have line["dir"] != (1, 0).
• But if we have page.rotation != 0 and all (or most) lines have line["dir"] = (1, 0) then we must not execute page.remove_rotation(). You can prevent this from happening already today, by simply setting page.set_rotation(0). Here si how you can implement it:

  blocks = page.get_text("dict")["blocks"]
  wdirs = [line["dir"] for block in blocks for line in block["lines"]]
  horizontals = [d for d in wdirs if d == (1, 0)]
  if len(horizontals) >= 0.9 * len(wdirs):  # 90% of all lines horizontal
      page.set_rotation(0)

I am considering a change in this sense.