7

I am using this binary square image of 15*15 pixels.

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

I am applying canny edge detection provided by openCV (version 2.7) for object size measurement. My expected output should look like,

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

But two edges (top and left edge) are always getting shifted by one pixel. The output of canny edge detection is,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 1 1 1 1 1 1 1 1 0 0 0
0 0 0 1 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 1 0 0 0 0 0 0 0 0 0 1 0 0
0 0 0 1 1 1 1 1 1 1 1 1 1 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Why is this pixel shift happening?
Is there any way I can avoid this. (I cannot manually adjust pixel shift after output, as I have to use edge detection on irregular shapes) The same shift happens irrespective of odd / even pixels.

Jeru Luke
  • 20,118
  • 13
  • 80
  • 87
Aditya Kanade
  • 71
  • 3
  • If you are using Python, can you try with [`skimage.feature.canny`](http://scikit-image.org/docs/dev/api/skimage.feature.html?highlight=canny#skimage.feature.canny) for comparison? – Catree Mar 24 '17 at 14:52
  • use findContours function instead – Micka Mar 24 '17 at 16:36
  • really really helpful question – Moritz Mar 24 '17 at 18:20
  • @Catree I have tried skimage canny. for the given ideal plane it is working good and the edges are not getting shifted. So I'll try it for other shapes. Thank you for your suggestion. – Aditya Kanade Mar 29 '17 at 10:39
  • @AdityaKanade From the `skimage` documentation, you can access to the corresponding [source code](https://github.com/scikit-image/scikit-image/blob/master/skimage/feature/_canny.py#L53) (here for the master branch). I would not be surprise if the result of OpenCV Canny() is not pixel perfect as the purpose of this library is also to get good performance (`findContours()` should return more accurate results). If you really want to spot the difference, I would suggest you to print every intermediate result in the source code of both versions and compare. – Catree Mar 29 '17 at 11:18

1 Answers1

6

At first glance, I was quite surprised when I came across this question. Moreover I did not believe that Canny edge detection would be so deceiving. So I took a similar image and applied Canny edge to it. To my surprise I encountered the same problem you are facing. Why is it so?

After digging in to the documentation I came across many operations that were occurring under the hood.

The documentation claims that Gaussian filtering is done to reduce noise. Well, it is true. But this blurs out the existing edges present in the image as well. So when you blur a perfect square/rectangle, it tends to have curved corners.

After Gaussian filtering, the next step is finding edge gradient. As said, by now the perfect edge of the square/rectangle is gone due to blurring (Gaussian filtering). What is left are rounded/curved edges. Finding the intensity of gradients on rounded/curved edges will never yield a perfect square/rectangle -like edge. I might be wrong, but I guess this the main reason as to why we do not get perfect edges while performing Canny edge detection.

If you want a perfect edge my suggestion would be to try finding contours(as suggested by Micka) and draw a bounding rectangle.

Jeru Luke
  • 20,118
  • 13
  • 80
  • 87