Image Clustering

JuSPICE provides a suite of unsupervised image clustering algorithms that partition an image into spatially coherent regions based on intensity, texture, or graph-theoretic criteria. These are distinct from the deep-learning and foundation-model pipelines described in Image Segmentation, which JuSPICE exposes through the separate spice.segmentation accessor. Pick an algorithm below.

Watershed

Distance-transform flooding that separates touching objects.

SLIC Superpixels

Compact, colour-homogeneous superpixels via local k-means.

Felzenszwalb

Efficient graph-based segmentation with an adaptive threshold.

Otsu Multi-Threshold

Histogram-based optimal thresholding into N intensity classes.

Quickshift

Kernel-density mode-seeking with a hierarchy of segmentations.

K-means

Per-pixel feature-vector clustering across every channel.

All six algorithms are accessible through ClusteringAccessor (spice.clustering.*).

Core API

See juspice.clustering_module for full API details and Clustering for a module-level explanation.

Notebook example

A complete workflow example is available in:

• notebooks/clustering/clustering_basics.ipynb (see Notebook Examples)

The notebook loads a SEM image via load_data(), applies all five algorithms, and saves the watershed label array through save_data() so that the result includes a provenance JSON sidecar and a reproducible history script.

For K-means clustering over multi-channel per-pixel feature vectors, see:

• notebooks/io/io_afm2.ipynb (see Notebook Examples) — clusters all 8 channels of an AFM .spm scan with spice.clustering.kmeans_clustering().

References

Libraries

• scikit-image — provides five segmentation algorithms (watershed, SLIC, Felzenszwalb, Otsu multi-threshold, Quickshift)

• scikit-learn — provides K-means clustering

• NumPy — label array storage and statistical post-processing

See Bibliography for the publications behind the methods on this page.