Georgios Balaouras
 Back to projects

Project February 2019

Image Segmentation

A comparison of spectral clustering and normalized cuts for graph-based image segmentation.

Code (opens in a new tab)
Comparison of a source image with spectral clustering and normalized-cut segmentation results.
Project image: Comparison of graph-based image segmentation methods. Open full-size image ↗

Project details

I represented an image as an undirected graph and grouped pixels into segments using similarities such as color and intensity. The project compares three approaches:

  • Spectral clustering: Use graph Laplacian eigenvectors to group pixels into a chosen number of clusters (method).
  • Normalized cuts: Partition the graph while accounting for the strength of connections within and between segments (paper).
  • Recursive normalized cuts: Repeatedly split segments using the normalized-cut criterion (paper).

The figure above shows how the resulting segments differ for the same source image.