Clustering

“Clustering” groups similar pixels together without needing any labeled examples — an unsupervised method. juspice.clustering_module gives you six clustering algorithms (watershed, SLIC, Felzenszwalb, multi-Otsu, Quickshift, K-means) through the spice.clustering accessor.

Native interface — ClusteringAccessor

spice.clustering is the main way to do unsupervised image clustering in JuSPICE. Each method returns a new SPICEData holding the resulting label array, and shares the input’s DatasetHistory — it never changes spice.data itself. This is the same derived-object pattern used by juspice.feature_extract.FeatureExtractionAccessor.extract():

ws_spice   = spice.clustering.watershed_clustering(footprint_size=60, min_distance=20)
slic_spice = spice.clustering.slic_clustering(n_segments=200)
fz_spice   = spice.clustering.felzenszwalb_clustering(scale=500)
otsu_spice = spice.clustering.otsu_multithreshold(n_classes=2)
qs_spice   = spice.clustering.quickshift_clustering(kernel_size=5)
km_spice   = spice.clustering.kmeans_clustering(n_clusters=3)

Available algorithms

  • watershed_clustering() — Marker-based watershed. Intermediate states (binary mask, distance transform, markers) returned in result.extra.

  • slic_clustering() — Simple Linear Iterative Clustering superpixels.

  • felzenszwalb_clustering() — Graph-based Felzenszwalb segmentation.

  • otsu_multithreshold() — Multi-level Otsu thresholding. Computed thresholds in result.extra and result.metadata.

  • quickshift_clustering() — Quickshift mode-seeking segmentation.

  • kmeans_clustering() — K-means clustering (scikit-learn) over per-pixel feature vectors. Fitted cluster centers and inertia returned in result.extra and result.metadata respectively.

Watershed, SLIC, Felzenszwalb, Otsu, and Quickshift all work on the same preprocessed grayscale image, in the [0, 1] range, kept ready for the whole session at spice.clustering.image. kmeans_clustering is the exception: it clusters the raw spice.data array directly, without collapsing it to grayscale first — so for a multi-channel dataset (e.g. all channels of an AFM .spm scan), every channel feeds into each pixel’s feature vector, not just one grayscale value.

Reproducibility

Each ClusteringAccessor() call records a clustering.<method_name> operation (e.g. clustering.watershed_clustering) into spice.history. to_lines() turns these back into plain, runnable code:

ws_spice   = spice.clustering.watershed_clustering(footprint_size=60)
slic_spice = spice.clustering.slic_clustering(n_segments=200)
km_spice   = spice.clustering.kmeans_clustering(n_clusters=3)

See also