Image To Graph¶

pleat.image_to_graph digitises a raster image of a line drawing (e.g. a scanned or photographed hand-drawn crease pattern) into a planar half-edge graph.

The pipeline is:

1. downsample the image,
2. threshold to a binary foreground mask,
3. close gaps with a morphological closing,
4. skeletonise and prune the result,
5. detect branch points and the edge segments between them,
6. merge branch points that are closer than a chosen cutoff,
7. emit an EuclideanPositionHEG.

The function plots intermediate results so you can pick threshold and edge_length_cutoff interactively.

Requires the optional dependencies scikit-image and mahotas.

In [1]:

import matplotlib
matplotlib.rcParams['figure.figsize'] = (5, 5)

import pleat
from pleat.image_to_graph import image_to_graph

import matplotlib matplotlib.rcParams['figure.figsize'] = (5, 5) import pleat from pleat.image_to_graph import image_to_graph

/home/runner/work/pleat/pleat/.venv/lib/python3.12/site-packages/mahotas/morph.py:315: SyntaxWarning: invalid escape sequence '\s'
  '''
/home/runner/work/pleat/pleat/.venv/lib/python3.12/site-packages/mahotas/features/texture.py:33: SyntaxWarning: invalid escape sequence '\|'
  '''
/home/runner/work/pleat/pleat/.venv/lib/python3.12/site-packages/mahotas/features/texture.py:158: SyntaxWarning: invalid escape sequence '\|'
  '''

Convert the image¶

We use the photograph test_graph_image.jpg shipped under misc_outputs/. The two parameters that almost always need tuning are:

• threshold — grayscale cutoff separating ink from paper. After the initial run the function shows you the grayscale histogram so you can pick a value.
• edge_length_cutoff — distance below which neighbouring branch-point clusters are merged into a single vertex. The function shows you the distribution of edge lengths to help pick this.

The function prints/plots the original image, the grayscale conversion, the thresholded foreground, the morphological closing, and finally the labelled edges.

In [2]:

G = image_to_graph(
    'images/test_graph_image.jpg',
    threshold=75,
    closing_iterations=3,
    edge_length_cutoff=40,
)

G = image_to_graph( 'images/test_graph_image.jpg', threshold=75, closing_iterations=3, edge_length_cutoff=40, )

Inspect the resulting graph¶

image_to_graph returns a fully-fledged EuclideanPositionHEG — the same data structure produced by the tile-growth pipeline. Lengths and angles have already been recomputed from the extracted vertex positions, so it is ready for downstream processing (Conway operators, reciprocal figures, folding, etc.).

In [3]:

print(f'{len(G.vertices)} vertices, {len(G.halfedges_representing_edges())} edges, {len(G.faces)} faces')
G.show()

print(f'{len(G.vertices)} vertices, {len(G.halfedges_representing_edges())} edges, {len(G.faces)} faces') G.show()

40 vertices, 98 edges, 59 faces

Saving and loading¶

Digitising a drawing is comparatively expensive, so you typically want to persist the result. Pleat's native YAML format .heg is round-trippable via pleat.io.save_graph and pleat.io.load_graph:

In [4]:

# pleat.io.save_graph('graphs/test_graph_image.heg', G)
# G_loaded = pleat.io.load_graph('graphs/test_graph_image.heg')
# G_loaded.show()

# pleat.io.save_graph('graphs/test_graph_image.heg', G) # G_loaded = pleat.io.load_graph('graphs/test_graph_image.heg') # G_loaded.show()