Unable to Extract Images from PDF with PyMuPDF and PyMuPDF4llm

[HanXiaoyou]

Description of the bug

Both PyMuPDF and PyMuPDF4llm fail to extract images from a specific PDF, despite the presence of images. No errors occur during extraction; the images simply aren't retrieved.

Versions:

PyMuPDF: 1.25.1
PyMuPDF4llm: 0.0.17
Additional Info:

PDF is not encrypted.
Tried different extraction methods, but the issue persists.

How to reproduce the bug

Both PyMuPDF and PyMuPDF4llm fail to extract images from a specific PDF, despite the presence of images. No errors occur during extraction; the images simply aren't retrieved.

Versions:

PyMuPDF: 1.25.1
PyMuPDF4llm: 0.0.17
Additional Info:

a.pdf

PDF is not encrypted.
Tried different extraction methods, but the issue persists.

PyMuPDF version

1.25.1

Operating system

Windows

Python version

3.9

[JorjMcKie]

This is no bug, but is caused by the weird way your PDF stores its images. Some background info for you:

• An image may or may not be identifiable by an xref. If it is not, then it is (probably) an "inline" image, which means it is stored in the page's /Contents ... side-by-side with other things like text.
• Inline images can be identified but not (yet) extracted by PyMuPDF (/4LLM). Identifying all images works via page.get_image_info(). If you use the xrefs=True parameter, then items will contain an xref which is positive for xref-ed images. xref=0 means "inline image".
• PyMuPDF4LLM uses the "bbox" key of get_image_info() items to get the display location on the page. It then makes a Pixmap of that bbox and includes this in the markdown output. In this way, a "photo" is made which includes inline images (and likewise vector graphics).
• PyMuPDF4LLM however, to prevent microscopic garbage, has a parameter-controlled limitation for image bbox sizes, parameter image_size_limit=0.05, which means that images are considered only if their bbox width and height are larger than then 5% of the page's width / height.
• The important issue of your file:
  • It only contains inline images(page 2).
  • But not 3 images as it seems: the number of inline images is around 340!
  • The PDF creator has cut down each original image to horizontal stripes with heights around 1 point (1/72 inches). Therefore, all 340 images are ignored because of the mentioned default of 5% (~40 points given a Letter page height).
  • There is no way to really heal this. You can of course set image_size_limit=1e-7 or similar, which will generate 340 images looking like more or less black/gray stripes. Probably not much of what you want.

[JorjMcKie]

You could of course look at page.get_image_info() and join the bboxes of the items you find in there.
Pay a little attention to join the appropriate ones of those stripes.
If done clever enough, you will end up with 3 larger bboxes.
For each of them do pix = page.get_pixmap(clip=bbox) which you can store in the usual way.