I know that S-function (membership function from fuzzy set theory) is a commonly used function for image enhancement. I am reading a paper titled - A Novel Fuzzy Image Enhancement Using S-Shaped Membership Function. The S-shaped membership function is defined as follows:

$$\begin{eqnarray*} S(x;a,b,c)=\begin{cases} 0, & x\leq a\\ \frac{(x-a)^2}{(b-a)(c-a)}, & a<x\leq b\\ 1-\frac{(x-a)^2}{(c-b)(c-a)} & b\leq x\leq c\\ 1, & x>c. \end{cases} \end{eqnarray*}$$

where the parameters are to be found. In the above mentioned paper, it given that the values of a, b, c are entropy(I), median(I) and (max(I)+mean(I))/2 respectively. But there's no explanation for how and why these values were chosen. I tried to look for it in many papers referred by the above paper but they are of no use. Can someone give me some idea please? Thank you in advance.

  • $\begingroup$ a more general term for this is Sigmoid function. and scaled the way yours is to go from 0 to 1, it's probably another version of a smoothstep function. $\endgroup$ – robert bristow-johnson Jan 18 '19 at 18:44
  • $\begingroup$ thanks @OlliNiemitalo for finding the pdf of the paper. the OP did copy Eq (4) over correctly, but i am suspicious that the definition is wrong because it does not appear to be continuous at $x=c$. when $x=c$, then $$ S(x) = 1 - \frac{c-a}{c-b} \ne 1 $$ when $x$ is slightly larger than $c$, then $S(x)=1$. $\endgroup$ – robert bristow-johnson Jan 18 '19 at 20:18
  • $\begingroup$ it maybe should be $$S(x;a,b,c)=\begin{cases} 0, & x\leq a\\ \frac{(x-a)^2}{(b-a)(c-a)}, & a<x\leq b\\ 1-\frac{(x-c)^2}{(c-b)(c-a)} & b\leq x\leq c\\ 1, & x>c. \end{cases}$$ i am not sure. to be a decent Sigmoid function, it should be continuous everywhere and also the first derivative should be continuous everywhere. $\endgroup$ – robert bristow-johnson Jan 18 '19 at 20:21

Okay, I am convinced that there is a typo in Eq. (4) of the cited paper. It should be:

$$ S(x;a,b,c)=\begin{cases} 0 & x\leq a\\ \frac{(x-a)^2}{(b-a)(c-a)} & a<x\leq b\\ 1-\frac{(x-c)^2}{(c-b)(c-a)} & b\leq x\leq c\\ 1 & x>c \end{cases} $$

This sigmoid function or smoothstep function is continuous everywhere and the first derivative is continuous everywhere, but the second derivative is discontinuous when $x=a$ or $x=b$ or $x=c$.

You must have $a < b < c$.

$a$ is defined where $S(x)$ begins to transition from 0 upward. and $c$ is defined for when the transition is over and $S(x)=1$. $b$, which is in between $a$ and $c$ is the point of inflection for when the curvature changes from a concave curvature to a convex curvature (or maybe it's the other way around).

There are better sigmoid or smoothstep functions than this that the author of the paper uses. I dunno why he/she is so enamored with this definition.

| improve this answer | |
  • $\begingroup$ Thank you so much for your response. Sir, I want how the parameters are determined. Can you help on this please? $\endgroup$ – shwetha Jan 20 '19 at 10:50
  • $\begingroup$ from what data would you want to fit the parameters? If $b$ is half-way between $a$ and $c$, then the curve is symmetric about the point of inflection. Where to put $a$ and $c$ depends on the values of $x$ where you want the sigmoid function to kick in and where on the $x$-axis you want the sigmoid function to finish and be flat again. $\endgroup$ – robert bristow-johnson Jan 21 '19 at 9:57
  • $\begingroup$ my exact question is, how do you decide that choosing such a parameter will enhance the given image? $\endgroup$ – shwetha Jan 23 '19 at 10:45

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.