# A mathematical way to represent an image kernel?

How to represent the calculation in this image mathematically? For example: With the discrete convolution and Fourier Transform.

It tries to do a calculation on the original image (image $$A$$/input) and save the result of the calculation in image $$B$$/output. It tries to multiply the pixel you are dealing with, with $$5$$ and it tries to multiply one pixel at the left, one pixel at the right, one pixel above of it and one pixel under it with $$-1$$ and adds the value of all the $$5$$ pixels you just multiplied together. After that, it will save the value of the new pixel in picture $$B$$. This happens to every single pixel in the image. Since there are no surrounding pixel on the corners, it simply takes them from somewhere else.

Which makes more sense?

Solution A:

$$c(i,j) = \displaystyle\sum_{k_1 \in \Bbb N} \displaystyle\sum_{k_2 \in \Bbb N} a(k_1, k_2) b(i - k_1, j - k_2)$$ where $$a$$ and $$b$$ are your matrices.

Supposing $$a$$ is your picture, by setting $$b(0,0)=5$$, $$b(±1,0)=−1$$, $$b(0,±1)=−1$$ and $$b=0$$ elsewhere.

Solution B:

$$(f \ast g)(x,y) = \sum_{i=-\infty}^\infty \sum_{j=-\infty}^\infty f(i,j) g(x-i,y-j).$$ $$f$$ is the filter, $$g$$ is the image, and $$f \ast g$$ is the filtered image. We have $$f(0,0) = 5$$, $$f(-1,0) = f(1,0) = f(0,-1) = f(0,1) = -1$$, and $$f(i,j) = 0$$ otherwise.

• nice animation interesting.+1 – dato datuashvili Apr 7 '13 at 9:11
• i think it is called Walsh-Hadamard Transform,is not it? – dato datuashvili Apr 7 '13 at 9:12
• en.wikipedia.org/wiki/Kernel_(image_processing) – dato datuashvili Apr 7 '13 at 9:16
• @dato That doesn't explain how te represent it with a discrete convolution? – user1095332 Apr 7 '13 at 9:17
• That website I linked to shows a mathematical description of the formula. You don't need to break it up into a series of two one dimensional convolutions. $$c(i,j) = \displaystyle\sum_{k_1 \in \Bbb N} \displaystyle\sum_{k_2 \in \Bbb N} a(k_1, k_2) b(i - k_1, j - k_2)$$ where $a$ and $b$ are your matrices. Supposing $a$ is your picture, by setting $b(0,0) = 5, b(\pm 1, 0) = -1$ and $b(0, \pm 1) = -1$ and $b = 0$ elsewhere, you will get the desired effect. – muzzlator Apr 7 '13 at 10:39

Let $A$ be an $m\times n$ matrix representing the input image, and $K$ be an $s\times t$ matrix representing the kernel, with $s$ and $t$ odd (since we need it to have a central entry). We will index $A$ with $$(i,j)\in \{1,\dotsc,m\}\times\{1,\dotsc,n\}$$ and we will index $K$ with $$(h,k)\in\left\{-\sigma, \dotsc, \sigma\right\}\times\left\{-\tau, \dotsc,\tau\right\}$$ where $\sigma=\left\lfloor \frac{s}{2}\right\rfloor$, $\tau=\left\lfloor \frac{t}{2}\right\rfloor$, and with $\lfloor \cdot \rfloor$ we mean the floor function. We do this so that the central entry of $K$ will have index $(h,k)=(0,0)$.
To take care of the edge entries we add two rows and two columns to $A$ by repeating the first and last row and column. That is, we define the extra entries \begin{align} A(0,0)&:=A(1,1) & A(0,j)&:=A(1,j) & A(0,n+1)&:=A(1,n)\\ A(i,0)&:=A(i,1) & & & A(i,n+1)&:=A(i,n) \\ A(m+1,0)&:=A(m+1,1) & A(m+1,j)&:=A(m,j) & A(m+1,n+1)&:=A(m,n) \end{align} This corresponds to the "extension" edge handling method. You can easily change those extra rows and columns to adapt to other methods.
Then for any fixed $(i,j)\in\{1,\dotsc,m\}\times\{1,\dotsc,n\}$ the entry $(i,j)$ of $K*A$ will be $$(K*A)(i,j) = \sum_{h=-\sigma}^{\sigma}\sum_{k=-\tau}^{\tau} K(h,k)\;A(i+h,j+k)$$