Spectral Segmentation Explorer

Build an affinity graph over image patches, take eigenvectors of the Laplacian, and cut the image where the graph is weakest.

  • spectral clustering
  • graph Laplacian
  • affinity
  • segmentation

Original patch

Spectral segmentation (2 regions)

Fiedler vector

Sign changes indicate the spectral cut boundary.

Smallest eigenvalues of the graph Laplacian

1234567891000.020.040.060.080.10.120.140.160.180.2Index (1 = smallest)Eigenvalue

A large gap suggests the natural number of clusters.

Graph weight matrix

σ_s=3 | σ_i=0.15

Graph construction is itself a model
Small σ_s produces local, fragmented cuts; large σ_s produces global, smooth cuts that may merge distinct anatomical regions. Spectral clustering adds spatial contiguity that intensity k-means lacks, but the weight function's parameters are modeling choices that must be validated, not read off the data.