1
$\begingroup$

I know there are different ways for partial derivation of an image, among others: Sobel kernel, LoG, Prewitt and so on.

But the simplest one is the central difference:

$$ \frac{d}{dx} f(x) \approx \frac{f(x+1) - f(x-1)}{2} \longrightarrow 0.5[1\ 0\ -1] $$

Which means convolving the image with above matrix.

Assume the image looks like this:

$$ I(x,y) = \left( \begin{matrix} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ \end{matrix} \right) \in \mathbb R^{8\times8} $$

Convolving this image with matrix $$G_x = 0.5\ [1\ 0\ -1] \in \mathbb R^{1 \times 3} $$ results in:

$$ \lvert I \ast G \rvert = \left( \begin{matrix} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0.5 & 0.5 & 0 & 0 & 0.5 & 0.5 & 0 \\ 0 & 0.5 & 0.5 & 0 & 0 & 0.5 & 0.5 & 0 \\ 0 & 0.5 & 0.5 & 0 & 0 & 0.5 & 0.5 & 0 \\ 0 & 0.5 & 0.5 & 0 & 0 & 0.5 & 0.5 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ \end{matrix} \right) $$

The cells containing $0.5$ are the edges of the image in $x$ direction.

Know assume we would extend our filter $G_x$:

$$ G_x = 0.5 \left( \begin{matrix} 1 & 0 & -1 \\ 1 & 0 & -1 \\ 1 & 0 & -1 \\ \end{matrix} \right) \in \mathbb R^{3 \times 3} $$

Now convolving this filter with our image results to:

$$ \lvert I \ast G \rvert = \left( \begin{matrix} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & 1 & 1 & 0 & 0 & 1 & 1 & 0 \\ 0 & 2 & 2 & 0 & 0 & 2 & 2 & 0 \\ 0 & 3 & 3 & 0 & 0 & 3 & 3 & 0 \\ 0 & 3 & 3 & 0 & 0 & 3 & 3 & 0 \\ 0 & 1 & 1 & 0 & 0 & 1 & 1 & 0 \\ 0 & 2 & 2 & 0 & 0 & 2 & 2 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ \end{matrix} \right) $$

Now instead of one unique number for edge like we had before with $0.5$ we get a gradient in $x$ direction $[1, 1, 2, 2, 3, 3, 2, 2, 1, 1]$.

Now my Questions:

1) Which approach is better, convolving the image with a $\mathbb R^{1 \times 3}$ -Filter or with a $\mathbb R^{3 \times 3}$-Filter?

2) And why is one better than the other?

Thanks

$\endgroup$
2
$\begingroup$

Non-zero pixels after derivation provide you with two informations: a potential location (where the pixel is), and a potential strength (in magnitude) of an edge-pixel. Linear edge detectors are linear filters. A $3\times3$ kernel adds a smoothing effect in the orthogonal direction. It will favor edges with a certain spatial extend, over short-lengthed isolated pixels that do not qualify for real-edges. In other words, an isolated $[0,1,0]^T$, a duplet $[1,1,0]^T$ will be affected a smaller magnitude than three vertical ones $[1,1,1]^T$ after a $3\times3$ kernel.

In different words, the Prewitt $3\times 3$ separable gradient/smoothing kernel can be written as a (Cartesian) product of two $3\times 1$ and $1\times 3$ kernels:

$$ \left( \begin{matrix} 1 & 0 & -1 \\ 1 & 0 & -1 \\ 1 & 0 & -1 \\ \end{matrix} \right) = \left( \begin{matrix} 1 \\ 1 \\ 1 \\ \end{matrix} \right) \times \left( \begin{matrix} 1 & 0 & -1 \\ \end{matrix} \right) $$

This scheme can be generalized: any smoothing filter in one direction, combined with any derivative in the other (orthogonal) direction, can form an horizontal or vertical square/rectangular image edge detector. You can find a source in Smoothed Differentiation Filters for Images, Meer and Weiss, 1992:

For the two-dimensional case we restrict ourselves to square neighborhoods. In such neighborhoods it is always possible to define a separable two-dimensional orthonormal basis built by the Cartesian product of two identical one-dimensional bases.

and in Derivative-based Operations or the wiki page Image derivatives .

If you have clear vertical edges, with same pixel amplitudes, little noise, the $1\times 3$ filter will do the job. Otherwise, good $3\times 3$ are almost always preferable. An image is really a 2D thing, and processing it through 1D tools often wastes information.

$\endgroup$
  • $\begingroup$ You can have some hints on other operators at dsp.stackexchange.com/questions/19175/… $\endgroup$ – Laurent Duval Jan 12 '17 at 10:27
  • 1
    $\begingroup$ @Laurant Duval could you give me some more references about this topic. I have already read some books about this topic but no one mentioned for example the smoothing effect in the orthogonal direction! $\endgroup$ – arash javan Jan 12 '17 at 11:04
  • $\begingroup$ I have added early references, tell me if you need more $\endgroup$ – Laurent Duval Jan 15 '17 at 13:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.