parameters and objective assessments for edge detection algorith

Message-ID:<aaa0788c-1913-4960-ab09-00353cb62ea8@a32g2000yqm.googlegroups.com>
Subject:

parameters and objective assessments for edge detection algorithms?

Date:Fri, 15 Jan 2010 16:35:16 +0100


Hi all,

I have some questions about edge detection algorithms. As far as I
know, now Canny is still one of the best algorithms for edge detection
(accurate, fast, and simple).

However, for Canny, and any other edge detections algorithms, we need
to choose some input parameters, for example for canny we need to
choose the high and low thresholds. And depending different images, we
will need different thresholds.

I am wondering if is there any method or paper discussing how to
automatically choose suitable parameters for an edge detection
algorithm (like Canny)? Please give any idea or a reference paper
regarding this issue.

The second question is that do you know any way to objectively assess
how well an edge detection algorithm does? (i.e. I mean a method that
we do not need some subjects to look at the results and give some
comments).

Thanks,

Message-ID:<cc7b00f3-b9fc-4889-b3ff-76988a3d105f@k17g2000vbm.googlegroups.com>
Subject:

Re: parameters and objective assessments for edge detection algorithms?

Date:Sat, 16 Jan 2010 06:27:58 +0100


I don't believe so.  ALL edge detectors are objective tools that
measure edges (although some have subjectively set parameters).  They
might differ slightly but they're all objective.  How can you say any
one is better than another?  There can be no "ground truth."  If there
were, what would it be?  It would be just another edge detector's
opinon of what the edge strength should be.  But who says that that
one is the absolute truth?  So all you can do is to compare the
results of the different edge detectors to themselves.  In fact, what
an edge actually IS, is somewhat inherently subjective, except in the
simple case of a sharp step function.  Let's say that you have this
profile
3 4 5 4 6 7 6 6 7 8 9 10 13 15 19 22 23 22 23 25 26 27 24
Okay, now where is the edge and what is its strength?  There is no
answer, and there are tons of answers.

So how "well" an edge detector works is inherently a subjective
opinion.  That's probably why there are hundreds of them.  You can go
here to get a list of them:

http://iris.usc.edu/Vision-Notes/bibliography/contentsedge.html#Edge%20Detection%20and%20Analysis,%20Lines,%20Segments,%20Curves,%20Corners,%20Hough%20Transform

There's even a section there for "optimal" edge detectors, but who is
to say what optimal or best is.  My optimal may be different than your
optimal.  In the end, optimal is what works best for you in your
situation with your specific images.  And thus we're back to
subjective human judgment.

Message-ID:<a126f4e4-365d-44c7-b7ae-4b0ad0ee7148@r19g2000yqb.googlegroups.com>
Subject:

Re: parameters and objective assessments for edge detection algorithms?

Date:Sat, 16 Jan 2010 14:19:36 +0100


ImageAnlyst,

Thank you for your reply. I guess you are answering my second
question :  there exists or doesn't exists an objective metric to
assess how well an edge detection performs.

You are right that the assessment depends on each person's opinion.
However, I think there may be some criteria which may be applied to
determine what edge detector is better than the other. For example, in
Canny's paper he listed out three or four (I do not remember) criteria
that a good edge detector should meet. I think it is similar to a non-
reference image quality assessment metric.

Thanks,





On Jan 15, 9:27=A0pm, ImageAnalyst <imageanal...@mailinator.com> wrote:
> I don't believe so. =A0ALL edge detectors are objective tools that
> measure edges (although some have subjectively set parameters). =A0They
> might differ slightly but they're all objective. =A0How can you say any
> one is better than another? =A0There can be no "ground truth." =A0If ther=
e
> were, what would it be? =A0It would be just another edge detector's
> opinon of what the edge strength should be. =A0But who says that that
> one is the absolute truth? =A0So all you can do is to compare the
> results of the different edge detectors to themselves. =A0In fact, what
> an edge actually IS, is somewhat inherently subjective, except in the
> simple case of a sharp step function. =A0Let's say that you have this
> profile
> 3 4 5 4 6 7 6 6 7 8 9 10 13 15 19 22 23 22 23 25 26 27 24
> Okay, now where is the edge and what is its strength? =A0There is no
> answer, and there are tons of answers.
>
> So how "well" an edge detector works is inherently a subjective
> opinion. =A0That's probably why there are hundreds of them. =A0You can go
> here to get a list of them:
>
> http://iris.usc.edu/Vision-Notes/bibliography/contentsedge.html#Edge%...
>
> There's even a section there for "optimal" edge detectors, but who is
> to say what optimal or best is. =A0My optimal may be different than your
> optimal. =A0In the end, optimal is what works best for you in your
> situation with your specific images. =A0And thus we're back to
> subjective human judgment.