fast-regionprops 0.1.0

Vectorized region-property features (skimage-compatible) via a single scatter pass.

fast-regionprops

A faster (vectorized) version of regionprops that supports a subset of region-property features, yielding a typical ~10–80× speedup (especially for many regions). Can be used as a drop-in replacement for skimage.measure.regionprops_table for the following set of properties:

label, area, centroid, bbox, equivalent_diameter_area,
intensity_mean, intensity_min, intensity_max, intensity_std,
inertia_tensor, border_dist

Dim	Image size	#Regions	skimage (ms)	fast (ms)	Speedup
2D	512×512	100	10.5	0.4	26.8×
2D	512×512	400	40.8	0.5	77.8×
2D	1024×1024	1600	163.1	2.0	82.8×
2D	2048×2048	3600	368.7	6.5	57.0×
3D	128×128×128	512	76.5	3.3	23.5×
3D	256×256×256	1728	229.6	19.6	11.7×

Install

pip install fast-regionprops
# dev / tests
pip install -e ".[test]"
# dev / tests / pre-commit
pip install -e ".[dev]"
pre-commit install

Usage

import numpy as np
from skimage.data import binary_blobs
from skimage.measure import label

from fast_regionprops import regionprops_table_fast

# example data: blobs as mask, random image as intensity
mask = label(binary_blobs(length=512, blob_size_fraction=0.05, volume_fraction=0.1))
intensity = np.random.default_rng(0).random(mask.shape).astype("float32")

# drop-in for skimage.measure.regionprops_table (same column names)
table = regionprops_table_fast(
    mask,
    intensity,
    properties=("label", "area", "centroid", "intensity_mean", "inertia_tensor"),
)
table["centroid-0"], table["inertia_tensor-0-1"]   # 1D arrays, one row per label

Output is a dict[str, np.ndarray] (same as for skimage.measure.regionprops_table)

Property notes

• Calling regionprops_table_fast(label_image) uses the same default properties as skimage.measure.regionprops_table: label and bbox.

• label_image must be an integer 2D or 3D label image. Boolean and floating point inputs are rejected as ambiguous; labels <= 0 are ignored.

• intensity_image may have the same shape as label_image, or one extra trailing channel axis. Multichannel intensity columns follow skimage naming, for example intensity_mean-0, intensity_mean-1.

• border_dist is an extra property (not in skimage): the per-label max of a field that fades from 1 at the image border to 0 once border_dist_cutoff px away (last two axes only).

• properties needing a convex hull / perimeter / topology (solidity, perimeter, euler_number, etc.) are not yet supported.

License

BSD-3-Clause.