- each pixel is its own cluster
- iteratively combine closest clusters until one remains
- result: dendrogram - tree of clusters
- lets you choose segmentation granularity “closest clusters”:
- single linkage → dist btw nearest pts in clusters
- long skinny clusters
- complete linkage → dist btw farthest pts in clusters
- tight clusters
- average linkage → mean dist
- (more robust to noise)
- inlier-outlier
pros & cons
✅ Adaptive cluster shapes (no fixed assumptions).
✅ No need to pre-specify cluster count.
❌ O(N²) runtime (slow for large images).
❌ Can get stuck in local optima (merges wrong clusters).