5
$\begingroup$

I try to resolve false edge problem in image segmentation. Let see its defination in Figure. The false edge problem often occurs in inhomogeneity image, especially for medical image. I am finding the method or feature to resolve it and achieve real edge. I try to apply some existing method example: localizing active contour, graph cut but they are not effective. I also apply gradient or entropy in here. Could you suggest some idea or feature or method to resolve it?Thank all . You can download test image at https://www.dropbox.com/sh/ttymp4vnpjhbvfx/2iacME5zYp . Note: orginal image is "input.mat". Its type is 16 bit

enter image description here

Let see example,In my marked image, B is real edge. We only analyse A point. At the A point, the intensity bigger than object region, but it not so big. Assume intensity of A point is 300, in which object intensity is about 260-280 (inhomogeneity property). So we can said A point is background, if we calssify it into two region. Sorry, you are true, the question must be update that A is true edge and C is false edge(because C is create by other region and background); true edge is created by object region and background

enter image description here

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ From this image it is not very clear what you atcually mean, perhaps you could clarify. Do you mean you are incorrectly identifying edges that do not relate to the object you want to detect and there are gaps in the "true edge"? If so I think this a very difficult problem to solve for the general case as there are loads of input parameters. For specific cases there are lots of approaches you could take, for example looking at edge strength, filtering the image or changing your edge detection threshold/method may help. $\endgroup$
    – nivag
    Commented Apr 24, 2014 at 13:19
  • $\begingroup$ Thank you for your support. I update my test image. You can down load and test it. The input is "input.dcm". The above figure only shows a simple defination of false edge problem. Let see my above figure, we can see in false edge region, the its intensity very similar with object region. So I think threshold is not effective idea in this case. Because my image is inhomogeneity image. Threshold is only effective in homogeneity image. Thanks $\endgroup$
    – John
    Commented Apr 24, 2014 at 13:39
  • 3
    $\begingroup$ I think this is a very hard problem to solve as there is little to no contrast along the line A (between the B & C regions). My best guess as you seem to know the segmentation you want would be to use some a priori knowledge to either reduce the "false" region or enhance the "true" region. My issue with such approaches as always a concern that it just tells you what you think should be there rather than what actually is. Sorry, I couldn't be more helpful. $\endgroup$
    – nivag
    Commented Apr 25, 2014 at 8:51

1 Answer 1

Reset to default
2
$\begingroup$

You could try "Non-maxima suppression" This is usually used on edge detectors to thin the edge to only one pixel. When there are lots of edges close together you basically pick the one with the largest gradient, and assume that is the true edge

Another technique that might be applicable is double thresholding. Some edge detectors (canny for instance) actually perform multiple detection's with different parameters. "Strong edges" are found in the the higher threshold (and probably the lower one as well) weaker edges may only be found using the lower threshold, and are only "true" edges if they connect to one of the stronger edges. Maybe you could use your current edge detector as the weaker classifier, and then make one that is harder to satisfy. This might eliminate the false positives that are only detected by the weak detector (i hope that makes sense)

good luck

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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