Scientific image processing on NumPy arrays · classical CV · verified against scikit-image 0.26 (2026)

scikit-image cheat sheet

scikit-image (skimage) is the scientific image-processing toolbox: filtering, thresholding, morphology, segmentation, feature detection, geometric transforms and region measurement — all on plain NumPy arrays, built on scipy.ndimage. It's the classical-CV complement to deep learning: preprocessing, quantification (cells, particles, materials) and analysis. The two rules to know: images are arrays (row, col) and dtype implies range (float in [0,1], uint8 in [0,255]). Targets scikit-image 0.26.

io, color, dtype filters & exposure morphology & segmentation measure & features transform & restoration gotcha ★ most common

Verified 2026-08-24 against the official docs at scikit-image.org (skimage 0.26). Builds on NumPy & SciPy; results are arrays you can hand to matplotlib, pandas (via regionprops_table), or an ML model.

Outline

A typical pipeline: read → grayscale/adjust → threshold/segment → label → measure regions. The submodules below map to those steps.

Load & Represent

Read images, and get the dtype/range right.

1io, sample data, dtype0.26

pip install scikit-image import skimage as ski # or: from skimage import io, filters, ...★
Submodules are imported explicitly (from skimage import filters) — there's no giant star-import.
img = ski.io.imread("cells.png") · ski.io.imsave("out.png", img)★
Returns a NumPy array (H, W) gray or (H, W, 3/4) color. as_gray=True to load grayscale directly.
img = ski.data.coins() # camera(), astronaut(), cells3d() ...★
Bundled sample images — perfect for trying pipelines and reproducing docs.
ski.util.img_as_float(img) · ski.util.img_as_ubyte(img)★
Convert between ranges safely: float in [0,1] vs uint8 in [0,255]. Never rescale by hand — these handle it correctly.

2Color & conversionsskimage.color

from skimage.color import rgb2gray, rgb2hsv, gray2rgb gray = rgb2gray(img)★
rgb2gray returns a float image in [0,1]. Many algorithms expect a single-channel float.
rgb2lab · rgb2hsv · lab2rgb · hsv2rgb
Full set of color-space round-trips for perceptual work (Lab for color distance, HSV for hue masking).
ski.color.label2rgb(labels, image=img)★
Colorize a label/segmentation map over the original — the standard way to visualize regions.

Enhance

Smooth, sharpen, find edges, fix contrast, and threshold.

3Filters & edgesskimage.filters

from skimage.filters import gaussian, sobel, median, unsharp_mask gaussian(img, sigma=2)★
Smoothing/denoise (gaussian, median), sharpening (unsharp_mask).
sobel(gray) · scharr(gray) · farid(gray)★
Gradient-magnitude edge maps. For crisp binary edges use feature.canny (see features card).
ski.filters.gabor(img, frequency=0.6) · ski.filters.meijering(...)
Texture (Gabor) and ridge/vessel filters (meijering/sato/frangi/hessian) for tubular structures.

4Exposure & thresholdingcontrast → binary

from skimage.exposure import equalize_adapthist, rescale_intensity, adjust_gamma equalize_adapthist(img) # CLAHE★
Contrast: CLAHE (local), equalize_hist (global), rescale_intensity (stretch), adjust_gamma.
from skimage.filters import threshold_otsu t = threshold_otsu(gray); binary = gray > t★
Otsu auto-picks a global threshold. Comparing to it gives a boolean mask — the start of most segmentation.
threshold_local(gray, block_size=35) · try_all_threshold(gray)★
Adaptive/local threshold for uneven lighting; try_all_threshold previews every method at once.

Segment

Clean up masks and partition the image into regions.

5Morphologyskimage.morphology

from skimage.morphology import binary_opening, binary_closing, disk binary_opening(mask, disk(3))★
Clean a binary mask: opening removes specks, closing fills holes. The footprint (disk/square) sets the scale.
erosion · dilation · ski.morphology.remove_small_objects(mask, 64)★
Shrink/grow regions; drop components below a pixel area (and remove_small_holes).
skeletonize(mask) · convex_hull_image(mask) · medial_axis(mask)
Shape analysis: 1-pixel skeletons, hulls, distance-based medial axis.

6Segmentationskimage.segmentation

from skimage.segmentation import slic, mark_boundaries segments = slic(img, n_segments=200, compactness=10)★
SLIC superpixels — group pixels into perceptually uniform blobs. mark_boundaries overlays them.
felzenszwalb(img) · quickshift(img) · chan_vese(gray)
Other unsupervised segmenters (graph-based, mode-seeking, level-set).
from skimage.segmentation import watershed labels = watershed(-distance, markers, mask=binary)★
Watershed splits touching objects (e.g. clumped cells) using distance-transform markers — the classic separation trick.
ski.segmentation.clear_border(mask)
Drop objects touching the image edge before counting/measuring.

Measure & Detect

Turn regions into numbers, and find keypoints/objects.

7Label & regionpropsquantify

from skimage.measure import label, regionprops, regionprops_table labels = label(binary) # connected components★
label assigns each connected blob an integer id — the bridge from mask to per-object measurement.
for r in regionprops(labels, intensity_image=img): r.area, r.centroid, r.eccentricity, r.mean_intensity★
Per-region properties: area, bbox, centroid, orientation, solidity, intensity stats — dozens available.
df = pd.DataFrame(regionprops_table(labels, img, properties=["area", "centroid", "mean_intensity"]))★
regionprops_table → a dict ready for pandas — the fast path to a measurements DataFrame.
ski.measure.find_contours(gray, level=0.5) · ski.measure.ransac(...)
Iso-value contours; robust model fitting (lines/circles) with RANSAC.

8Features & detectionskimage.feature

from skimage.feature import canny edges = canny(gray, sigma=2)★
The go-to crisp edge detector (hysteresis thresholding).
blob_dog(img) · blob_log(img) · peak_local_max(response)★
Blob/spot detection (cells, particles) and local-maxima finding (for watershed markers).
hog(img) · local_binary_pattern(gray, P, R) · corner_harris(gray)
Descriptors/textures (HOG, LBP) and corner/keypoint detectors (Harris, ORB).
match_template(img, template)
Normalized cross-correlation template matching — find a small patch in a larger image.

Transform & Restore

Resize/warp geometry and denoise.

9Geometric & restorationtransform / restoration

from skimage.transform import resize, rescale, rotate resize(img, (256, 256), anti_aliasing=True)★
Keep anti_aliasing=True when downscaling to avoid moiré. rescale by factor; rotate by degrees.
from skimage.transform import AffineTransform, warp warp(img, AffineTransform(scale=..., rotation=..., translation=...).inverse)★
Arbitrary geometric transforms (affine/projective/similarity). estimate_transform fits one from point pairs (image registration).
hough_line(edges) · hough_circle(edges, radii) · probabilistic_hough_line(...)
Detect lines/circles from an edge map.
from skimage.restoration import denoise_tv_chambolle, denoise_bilateral, inpaint
Edge-preserving denoising (TV, bilateral, wavelet) and region inpainting.

!Common gotchasread before shipping

dtype = value range
float images are assumed [0,1], uint8 [0,255]. Feeding a 0–255 float where [0,1] is expected silently breaks results — use img_as_float/img_as_ubyte.
coordinates are (row, col)
Points/centroids are (y, x), not (x, y) — the opposite of most plotting/OpenCV conventions.
many functions want grayscale float
Thresholds/filters/features often expect a single-channel float — rgb2gray first (it also returns float).
anti_aliasing when downscaling
resize/rescale default it on for you now, but confirm — off produces aliasing artifacts.
label before regionprops
regionprops needs an integer label image (from measure.label), not a boolean mask.

Worth memorizing

images are NumPy arrays(row, col[, channel]); submodules imported explicitly

dtype implies rangefloat [0,1], uint8 [0,255]; convert with img_as_float/ubyte

coordinates are (y, x)row, col — reversed vs OpenCV/plotting

threshold_otsu → maskgray > threshold_otsu(gray) starts most segmentation

morphology opening/closingremove specks / fill holes; footprint sets the scale

watershed splits touching objectsdistance transform + markers

label then regionpropsconnected components → per-object area/centroid/intensity

regionprops_table → DataFramethe fast path to measurements in pandas

resize with anti_aliasingavoids moiré when downscaling