Image Segmentation through Spectral Clustering

Jiwei Peng (McMaster University)

Monday, October 6, 11:30 (HH312)

Clustering is the act of grouping data of similar structures or behaviors. Any finite dataset can be transformed into an undirected, weighted simple graph by representing each data point as a vertex and assigning a weighted edge between every pair of vertices to reflect their degree of similarity. By analyzing the matrix representation of this similarity graph through the lenses of spectral graph theory and numerical linear algebra, one can determine an optimal partitioning of the graph which maximizes the similarity within subgraphs while ensuring the dissimilarity between them. This technique is known as spectral clustering. We provide a practical application of this method in image segmentation, where distinct objects within an image can be detected and extracted.