Trainable weka segmentation for electrode material

I have processed image using TWS, and attached are the sample image and processed image. The pores are not clear in the image and the brighter phase also getting dilated. Can you please suggestion which 2-d training feature I should select to preserve edges and clear segmentation of my sample image? What should be the value of sigma and why?
And can you please suggest where I can read more about these training feature selection criteria or bit more detailed description about them besides ImageJ help page?
I am struggling on this from a while, any help will be a great relief.

Thank you

Blue-white one is sample image
and the red-green one is output image

Hello @sam1990 and welcome to the ImageJ forum!

Excuse my ignorance about this type of images but what are the edges in the image? The dark areas? Could you point out in your result image which are the mistakes in the segmentation you found?

1 Like

Hi @iarganda
thanks for replying.
Yeah, the black one are pores in this material. White and blue are two kind of materials. Blue dots on white bits is just noise or signal from beneath material, so should be ignored. And , if we compare the sample image with segmented image, it seems pores (black in sample image)are not segmented clearly there and white particles have become dilated. So I wanted to preserve these edges of white and blue particles while segmenting.

I see. So what about the following image as result?

My doubts are in regions like this one:

Do you want the plugin to trace a border there?


Yes, this kind of segmentation I am looking for, what steps do you follow and which training feature you selected ?
And, then this particle boundary doubt like as you have asked. At location 2 there is boundary, these are two disconnected particles (just a disconnect to separate particle not visible pore though) and location 1 its connected. so how to deal with this also.

Thank you very much for your time and suggestions.

Thank you

I am curious what training features you are using, I am using Kuwahara and Anisotropic diffusion with the default ones (all five except membrane projection) and default values(1,19,1,16) ?

Since the image is basically “bluescale” I assumed the color information was not relevant so I converted it 8-bit:

Then I used the following features and parameters with just a few samples:

I assume you can get better results if you continue adding more representative traces/samples.

1 Like

Thank you.
And, how can i get the separate boundaries between particles ? the same question as you asked before, how to deal with that ?

Well, for me the first thing to try is to continue training with more examples of those boundaries:

What about this result?

Thanks, this seems the best result close to sample image, how did you achieve that? as training features as before ?

Yes, everything the same. I just kept tracing more examples of the conflictive areas :wink:

1 Like

Thank you, it was a great help.

1 Like

My pleasure!

Let me know if you need more help or if you want me to send you the classifier I trained or the ARFF with the data I used.

1 Like

Thank you very much.
It will be great if you can email those files. I have sent you a help request e-mail before, there you can send it.

You have saved me from a lot of confusions, now I can use this set procedure. Earlier I was just trying different training features and it was confusing.

Thanks a lot.

1 Like