How to extract and detect repeated texture patterns in a vector-based PDF?

[roy-projects]

Hi,

I have a vector-based architectural floor plan PDF that uses repeating textures (known as Hatches) to represent things like new walls or demolition. These patterns/textures also appear in a legend box.

I’m trying to:
1. Extract a texture from the legend area (Currently manually cropping it and saving as PNG).
2. Detect where it appears elsewhere on the page
3. Measure its total area.

I’d like to work with the vector data directly using PyMuPDF, rather than rendering to an image.

Is it possible to access and compare these texture fills or path objects in PyMuPDF? Any tips for matching repeated textures across regions?

Unfortunatly I am unable to share the original PDF file, but I did extact drawing and re-drawed them by layers to seperate PDF files that I can share.

Example of layer 1:

• This PDFs has the same layout as the original floor plan stracture - with a pattern on the left (Legend area) and the pattern re-drawn on the right (Engineering drawing area).
  500-Harisa.pdf

Example of layer 2:

• This PDF only has the engineering drawing area with 2 different patterns, while the legend patterns are un-organizingly located on a different pattern.
  Wall-new.pdf

I would highly appreciate any help with this situation!

Besides that, for a different task - symbols detection - I already used cluster drawing extraction, re-drawed each to a shape, and got good results with template matching (using cv2) so I wanted to say a big thanks, it's been a journy so far and this library is amazing and very educating!

[JorjMcKie]

Thanks for your appreciation and compliments.

Otherwise ... hm! It seems the issue is not the extraction itself but how to define & detect a new class of drawings.
Roughly similar (more elementary though) has been done in our table detection when we expanded the notion "line" somewhat to also include thin filled rectangles or multiple small line segments or little circles drawn in a dashing manner.
Without claiming to see a solution, maybe something similar is possible here. When doing .cluster_drawings() these objects should be contained in rectangular areas with "suspicious" aspect ratios (one of width:height or height:width is very small) and otherwise only containing lots of parallel line segments.

[JorjMcKie]

For example in 500-Harisa.pdf the first shape has a large width compared to a small height.
Its paths exclusively consist of small stroked lines and (almost) every line has a rectangle hull with the same height as the shape. The counts of y values are

pp(y_values)
{569.4000244140625: 212,
 563.6400146484375: 212,
 565.5599975585938: 2,
 568.5599975585938: 2,
 563.760009765625: 2,
 566.8800048828125: 2}

The first two values equal the shape's top and bottom coordinates ...