[WIP]Add skeletonization function

[Deepank308]

Added skeletonization algorithm using medial-axis transform for binary images.
I will add documentation soon.

Please give suggestions for improving the API and performance, if any.
Ping @juliohm
Thank you!

[johnnychen94]

I left some first-round comments here, it's just something I noticed, and I think it's more appropriate to let others review your code, they're more experienced and more qualified.
I didn't read the MedialAxis Skeletonization algorithm before, so I trust that you've written the algorithm procedure correct. Or could you give a reference to your implementation, so that I can have a look at? (I think it's quite an easy algorithm)

---

I haven't' reviewed others' codes before, so if there's anything I did not properly, please point it out :)

[johnnychen94]

There's using ImageMorphology in Images.jl, and it will make package dependency complicate.

@timholy How do we deal with this problem in general? Can we directly copy Images.FeatureTransform module to ImageMorphology as temporary not-exported functions?
https:/JuliaImages/Images.jl/blob/bdfd044420fa6ffcd34760f804f0c3ce12186945/src/bwdist.jl

[juliohm]

@johnnychen94 another example where we find something implemented in Images.jl that could live somewhere else. We really need to reorganize concepts around and break Images.jl into smaller packages for reuse. Right now a bunch of functionality lives in the umbrella package, and that is suboptimal.

[johnnychen94]

Looks like @Deepank308 needs this function implemented and merged ASAP to continue his GSoC ImageTracking project.

@juliohm I guess we can copy these methods to FeatureTransform.jl (with comments) and remove it in the future, just like @zygmuntszpak does in https:/zygmuntszpak/ImageBinarization.jl/blob/6e0c81867eaef67b463fa5d52febfca4d0f6196a/src/integral_image.jl ?

[juliohm]

I don't know if copying/pasting helps @johnnychen94 , ideally we would start thinking more seriously about how to reorganize things. There is a lot of code living in the wrong place and we are just compromising the situation further by adopting multiple copies in submodules.

[johnnychen94]

I might need another two weeks to start the porting work. Let's see how it works then.

[Deepank308]

@johnnychen94 I don't need this function as of now in my implementations. I was just going through Skeletonization algorithm and found that Julia doesn't have it. So, I implemented and sent a PR.

[johnnychen94]

In this case, this PR might be pended until FeatureTransform being ported to a standalone module.

[codecov]

Codecov Report

Merging #13 into master will decrease coverage by 29.24%.
The diff coverage is 72.09%.

@@             Coverage Diff             @@
##           master      #13       +/-   ##
===========================================
- Coverage     100%   70.75%   -29.25%     
===========================================
  Files           2        4        +2     
  Lines          44      106       +62     
===========================================
+ Hits           44       75       +31     
- Misses          0       31       +31

Impacted Files	Coverage Δ
src/ImageMorphology.jl	100% <100%> (ø)
src/skeletonization.jl	71.42% <71.42%> (ø)
src/thinning.jl	57.57% <0%> (-42.43%)	⬇️
src/dilation_and_erosion.jl	83.33% <0%> (-16.67%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 43b3d0e...d4f8638. Read the comment docs.

[codecov]

Codecov Report

Merging #13 into master will decrease coverage by 28.84%.
The diff coverage is 73.17%.

@@             Coverage Diff             @@
##           master      #13       +/-   ##
===========================================
- Coverage     100%   71.15%   -28.85%     
===========================================
  Files           3        4        +1     
  Lines          62      104       +42     
===========================================
+ Hits           62       74       +12     
- Misses          0       30       +30

Impacted Files	Coverage Δ
src/ImageMorphology.jl	100% <100%> (ø)	⬆️
src/skeletonization.jl	72.5% <72.5%> (ø)
src/thinning.jl	57.57% <0%> (-42.43%)	⬇️
src/dilation_and_erosion.jl	83.33% <0%> (-16.67%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 277925b...c3ef6bb. Read the comment docs.

[johnnychen94]

add a title to the reference item for clarity

also here are some thoughts on simplifying documentation of function with multiple methods.

JuliaImages/Images.jl#790 (comment)
JuliaImages/ImageBinarization.jl#25

It'll be great to hear thoughts from you.

Further documentation work can be done after codes being stable.

[Deepank308]

Noted. I will keep that in my mind if any other skeletonization algorithm will be implemented in future?

[Deepank308]

I don't know the exact reason as to why AppVeyor test is failing. The test fails for julia-latest.
It is showing error :
ERROR: UndefVarError: versioninfo not defined
Any ideas?

[johnnychen94]

An upgrade on CI configuration is needed, #14

As for ERROR: UndefVarError: versioninfo not defined error, we need using InteractiveUtils; versioninfo()

[Deepank308]

I was curious to compare the performance of the algorithm in Julia and Python(skimage). This is what I got :
I have compared for white rectangles over black background.

Image Size	Rect dims	Julia(in ms)	Python(in ms)
(300, 300)	(200, 200)	99ms	99ms
(300, 300)	(100, 100)	44ms	97ms
(600, 600)	(400, 400)	400ms	180ms

So, there is some serious need for performance improvements!! Please suggest if anyone has some ideas.
Thanks!

[johnnychen94]

according to my experiences:

• decreasing unnecessary memory reallocation can always improve performance for most Julia codes -- use @btime from BenchmarkTools to check it -- that's why we use StaticArrays
• check if the for-loops in corner_table_lookup and inner_skeleton_loop can be vectorized
• the append! in compute_critical_indices changes memory in every iteration, pre-initializing with zeros(Bool, length(critical_index_array)) and cutting the tail after the for-loop might help.

not sure if any of it works, sorry that I don't have much time digging into the details at present. There might be more tips in performance tweaking.

[fweth]

I wrote a simple grassfire transform algorithm which works also on non-binary images: https:/fweth/Julia/blob/master/grassfireTransform.jl

I experimented with rasterized vector graphics and found out that treating pixels with values < 1 as already displaying the 'true' distance to the border gives smoother results than first applying threshold to the rasterization.