譜圖(Spectral Graph Theory)理解(2)
阿新 • • 發佈:2018-12-10
參考文章:Introduction to Spectral Graph Theory and Graph Clustering 作者:Chengming Jiang,ECS 231 Spring 2016 University of California, Davis 本文的目的是進行計算機影象分割: 圖1 影象分割
一、預備知識
關於圖(G)、度矩陣(D)、鄰接矩陣(A)皆在上一篇理解中交代過,現補充一些新的定義: 1、權重矩陣 A weighted graph is a pair G=(V,W) where
- is a set of vertices and ;
- is called weight matrix with W是權重矩陣,V是頂點,它們構成對G=(V,W),即是權重圖G。 The underlying graph of G is with
- If , the adjacency matrix of
- Since , there is no self-loops in
W是對A的一個擴充套件,當,W即是A。定義W後,需要重新定義頂點的度(degree of a vertex)和度矩陣(degree matrix):
2、A的體積(Volume)
對於V的一個子集A(),定義A的體積(Volume):
即A中所有頂點的度和,若A中所有頂點都是孤立的(isolated),則vol(A)=0,舉例如下:
圖2 vol(A)的計算方法
3、頂點集間的連線(links)
Given two subsets of vertices , we define the links by
Remarks:
- A and B are not necessarily distinct;
- Since W is symmetric,
- 有了連線(links)定義,就可以定義分割(cut),它的定義如下: 在連線(links)基礎上,還可以定義一個量assoc,如下: 即A中頂點自己的連線。cut是A和外部的links,assoc是A與內部的links。因此有: 4、Graph Laplacian 對於權重圖 G=(V,W),the (graph) Laplacian L of G is defined by Laplacian具有以下的屬性:
- for ,這是一個二次型
- if for all i,j;
- If the underlying graph of G is connected, then
- If the underlying graph of G is connected, then the dimension of the nullspace of L is 1.
圖的聚類(Graph clustering)
1、k-way partitioning 給定一個權重圖 G=(V,W),要找到一個對V的分割,使以下條件得到滿足:
- for any i and j, the edges between have low weight and the edges within have high weight. 要使分割後各子集之間的edges的權重最小,對於2-way分割有: , where