Deduplicating tables and images

[jamesbraza]

With pymupdf==1.26.3 I have discovered that:

Source	Saved Pixel Map
Image found via page.get_drawings + page.cluster_drawings
Table found via page.find_tables

These are both roughly the same PDF content, but the table form is a bit "tighter".

Is there some robust PyMuPDF approach to deduplicate tables and images?

An example of what I'm looking for is some PyMuPDF built-in capability such as page.cluster_drawings having a boolean flag deduplicate_from_tables.

---

I have seen cases where:

• page.get_drawings + page.cluster_drawings is a large superset of page.find_tables, so a simple deduplication such as 90% bounding box overlap isn't robust
• page.find_tables gives garbage so I am afraid to prioritize tables over images

[JorjMcKie]

The table module already uses vectors and cluster_drawings to help detect tables.

[jamesbraza]

Okay gotchu, so the two routes have more commonality than I realized.

That being said, do you have any tips on deduplicating the two? Or is there not really a general heuristic here

[JorjMcKie]

Again, the PyMuPDF4LLM module does exactly this:

1. Create paths = page.get_drawings().
2. Extract tables reusing the paths: tabs = page.find_tables(paths=paths, ...). Will prevent extracting paths again by the Finder.
3. Select the subset of paths (call it e.g. no_tab_paths) that are not contained in any table bbox.
4. Find vector graphic clusters passing the sub paths to it: clusters = page.cluster_drawings(drawings=no_tab_paths). This will keep the method from doing path extraction itself.