Equality related to normalized graph cuts for image segmentation

61 Views Asked by At

Reading an image processing (image segmentation paper) entitled Normalized Cuts and Image Segmentation, the goal is to create a partition of a weighted graph $V$ into two sets $A,B$ such that $Ncut(A,B)$ is minimal.

Let $x$ be an $N=|V|$ dimensional indicator vector $x_i = 1$ if $i \in A$, and $-1$ otherwise. Let $d(i) = \sum_{j} w(i,j)$ be the total connection from node $i$ to all other nodes.

Define $Ncut(A,B) = \frac {cut(A,B)}{assoc(A,V)} + \frac {cut(B,A)}{assoc(B,V)}$ where $cut(A,B)$ is the weight of all edges connecting sets $A,B$, and $assoc(A,V)$ is the weight of all edges from $A$ to all nodes in the graph.

Let $D$ be an $N \times N$ diagonal matrix $[D_{ii}] = d(i)$, and let $W$ be an $N \times N$ symmetrical matrix $W(i,j) = w_{ij}$ giving the weights of our graph. Further let $k = \frac{\sum_{x_i > 0} d_i} {\sum_{i} d_i}$ (the sum of weights coming out from elements in $A$ normalized by the overall sum of weights leaving all nodes).

Writing $Ncut(A,B) = \frac {cut(A,B)}{assoc(A,V)} + \frac {cut(B,A)}{assoc(B,V)}$ $ = \frac{\sum_{x_i > 0, x_j < 0}-w_{ij}x_ix_j}{\sum_{x_i > 0} d_i} + \frac{\sum_{x_i < 0, x_j > 0}-w_{ij}x_ix_j}{\sum_{x_i < 0} d_i}$.

The authors state without proof that $4 \times Ncut(x) = \frac {(1+x)^T (D-W)(1+x)}{k1^TD1} + \frac {(1-x)^T(D-W)(1-x)}{(1-k)1^TD1}$, noting that $\frac{1+x}{2}$ and $\frac{1-x}{2}$ are indicator vectors for $x_i>0$ and $x_i<0$ respectively.

I'm a bit confused how this formula was arrived at. I tried coming up with my own example to see why this was true for the first part of the addition on the RHS. Imagining for a graph with two nodes, $x=(x_1,x_2)$, then $\frac{(1+x)^T(D-W)(1+x)}{k1^TD1} = \frac{((1,1)+(x_1,x_2))^T (\begin{bmatrix} d(1)&0 \\0 & d(2) \end{bmatrix} - \begin{bmatrix} w_{11} & w_{12} \\ w_{21} & w_{22} \end{bmatrix}) ((1,1)+(x_1,x_2))} { \frac {\sum_{x_i>0}d_i} {\sum_{i}d_i} (1, 1)^T \begin{bmatrix} d(1)&0 \\0 & d(2) \end{bmatrix} (1, 1)}$,

I can see the denominator equals $\sum_{x_i>0}d_i$, but I'm not seeing the relationship with the numerator between the two expressions.

1

There are 1 best solutions below

1
On BEST ANSWER

I'm not seeing the relationship with the numerator between the two expressions.

For instance, (assuming that $w(i,j)$ is the same as $w_{ij}$ and $D-W=\|(D-W)_{ij}\|$), the numerator of $\frac {(1+x)^T (D-W)(1+x)}{k1^TD1}$ equals $$\sum_{i,j}(1+x_i)(D-W)_{ij}(1+x_j)=$$ $$\sum_{i}(1+x_i)^2d(i)-\sum_{i,j}(1+x_i) (1+x_j)w_{i,j}=$$ $$\sum_{i,j}(1+x_i)^2w_{i,j}-\sum_{i,j}(1+x_i) (1+x_j)w_{i,j}=$$ $$\sum_{i,j}(1+x_i)(x_i-x_j)w_{i,j}=$$ $$\sum_{x_i>0,\,x_j<0}4w_{i,j}=$$ $$\sum_{x_i>0,\,x_j<0}-4x_ix_jw_{i,j}.$$