Re: Hysteresis step in Canny edge detection

Subject: Re: Hysteresis step in Canny edge detection
From: Olzhas Zhumabek (anonymous.from.applecity_at_[hidden])
Date: 2019-03-17 05:20:06

Thanks for heads-up.

I had a look at it, and it seems like it has the same problem as I had in
the beginning: applying hysteresis as I go, not at once in the end. From
what I understand, hysteresis makes transition from weak edge to strong
edge only when strong edge is in immediate vicinity, e.g. one of the
closest 8 pixels. The way it is written in the code linked it seems like it
will produce different outputs on these inputs (assuming they are somewhere
in the middle of image and immediate vicinity is filled with zeroes):

1, 0.5, 0.5,      ->   1, 1, 1
0.5, 0.5, 0.5           1, 1, 1

should be in my opinion
1, 1, 0.5,
1, 1, 0.5

-----------------------

0.5, 0.5, 0.5, -> 0, 0, 0
0.5, 0.5, 1            0, 0, 0

should be in my opinion
0.5, 1, 1,
0.5, 1, 1

E.g. relative position of weak edges to strong edges matters in code of the
linked pull request. I wanted to verify if I'm right before contacting the
author, because I've seen various interpretations on the internet.

I uploaded my solution to github:
https://github.com/simmplecoder/edge-detection . I separated IO code from
the rest, as it seems to cause massive compilation time increase. The toy
program asks for input file, output file, upper and lower threshold in
double thresholding. It saves all of the intermediate steps in .png files
as well (can be suppressed by defining LEAN_AND_MEAN at the top of main.cpp
file). The code has comments and more or less meaningful names, but might
be a little bit verbose. I'm paranoid on Argument Dependent Lookup, so I
always prefix with namespace if I actually don't want ADL. CMake script
assumes presence of GIL and Boost.Math for value of Pi in the include path
(last time I tried it seems like CMake doesn't know about existence of GIL).

Here is the output with thresholds 0.7 and 0.1: https://imgur.com/QcGxwzg .
Original image:
http://justin-liang.com/tutorials/canny/images/1_original.jpg . It looks
like due to difference in non-maximum suppression algorithm, mine picks up
quite a bit of false edges along the way.

What steps next should I take in my application process to Boost.GIL GSoC?

Sincerely,
Olzhas

> FYI, someone has just proposed the Canny edge detector,
> have a look at the proposed implementation at
> https://github.com/boostorg/gil/pull/258
>
> Best regards,
> --
> Mateusz Loskot, http://mateusz.loskot.net