GeoDict Forum
Image Processing and Image Analysis and Technologies => CT, µCT and FIB-SEM => Topic started by: jdrunner on June 11, 2025, 04:00:50 PM
-
We are trying to obtain fiber length and count through the FiberFind (AI) module.
The fibers are multidirectional and have fairly high surface roughness.
The module is currently identifying what should be single fibers as multiple fibers, and some fibers are not getting identified at all. The threshold is currently set to 0.9
If we increase the threshold, the issue of multiple fibers attributed to 1 decreases, but so does the identification of fibers. The below image captures what we are currently dealing with.
[attachimg=1]
-
Hello Jack,
Thank you for reaching out.
In case of an over-segmentation we recommend reducing the threshold, as described on page 40 and 55ff. of the FiberFind user guide https://www.math2market.com/fileadmin/UserGuide/GeoDict2025/FiberFind2025.pdf.
First, you can check for a sufficient threshold by analyzing the result you already have. To do this, you can load the numpy field (as described on page 55) and find a threshold value with which the centerlines are displayed properly.
Please also note that the diameter of the fibers have to be at least 8 voxels to get an stable identification result.
I hope this helps you with your work.
Best regards,
Lilli
-
Hi Lilli,
Thank you for your response.
We have tried different thresholding values, still with no luck.
We also dilated our structure to ensure that the diameter was at minimum 8 voxels to eliminate any instability there.
Our fibers have relatively high surface roughness, which might be playing a role in the fibers being broken up into multiple identified fibers.
I would be glad to send our structure file to you if that would help any.
Respectfully,
Jack Davis
-
Hello Jack,
yes, the noise might be also a problem.
There are some Expert Setting, which you can use in that case. Please include the attached .gps-file into the Settings > Edit Expert Settings... dialog before running the identification process once again.
Best regards,
Lilli
-
Hello,
After applying the filter, it seemed to make the identification worse, as the number of identified fibers increased.
I tried messing around with the threshold to see if that would get accurate results, but unfortunately it did not work.
Is there anything else that could help in identifying the correct number of fibers and their lengths?
Respectfully,
Jack Davis
-
Addendum:
We applied the filter, tried the threshold analysis, as well as dilation of fibers. It seemed to do slightly better but still broke up the fibers quite a bit.
Are there any other suggestions to enhance the results? We have exhausted what we currently know with no significant improvement, and our deadline is rather quickly approaching. Thank you.
Respectfully,
Jack Davis
-
Dear Jack Davis,
I'm sorry for the late reply. I think it would be great if you could send the structure file. If you don't want to share it here, send it to support@math2market.de. Then, we can have a look at it.
In case the current networks cannot handle this structure, it could help if you train your own custom network using GeoDict-AI specialized on your structures.
https://geodict-userguide.math2market.de/2025/geodictai.html (https://geodict-userguide.math2market.de/2025/geodictai.html)
There is also a tutorial for training a neural network for identifying binder. I will send you the password for unzipping the tutorial via e-mail.
https://www.gddownload.de/Tutorials/Training-Neural-Network-for-binder-identification-in-NMC-cathodes-with-GeoDictAI.zip (https://www.gddownload.de/Tutorials/Training-Neural-Network-for-binder-identification-in-NMC-cathodes-with-GeoDictAI.zip)
If you want to train your own network, I think the training data generation macro should then (at least) contain the following steps:
* use FiberGeo > Create to create the fibers https://geodict-userguide.math2market.de/2025/fibergeo_create.html (https://geodict-userguide.math2market.de/2025/fibergeo_create.html)
* create the fibers as close as possible to the original regarding fiber curvature, diameters, etc.
* save the structure as *.gad file (File > Save Structure as)
* use GrainGeo > Roughen Surface to roughen the surface https://geodict-userguide.math2market.de/2025/graingeo_roughensurface.html (https://geodict-userguide.math2market.de/2025/graingeo_roughensurface.html)
* use GadGeo > Add & Import > Add GAD Objects from File to add the previously saved gad file to the current structure to have the gad data and the roughened data in memory at the same time https://geodict-userguide.math2market.de/2025/gadgeo_addgadfile.html (https://geodict-userguide.math2market.de/2025/gadgeo_addgadfile.html)
Best regards,
Janine
-
Hi Jack Davis,
Good news: using the following expert settings (Settings > Edit Expert Settings) and a threshold of 0.7 we could improve the result significantly:
[attachimg=1]
Short explanation for the chosen expert settings :
- Reconnection Threshold: we strongly relaxed it because the fibers are very isolated and it does not connect through the pore space.
- Max Distance: this parameter is important because otherwise only the local fiber radius is considered.
- Remove Crossing Radius: this removes crossing artefacts at crossings, preventing branched fibers
- Cleanse Centerlines: this cleans the result improving the centerlines. This is needed because the fibers have such a rough surface. For this reason also a higher threshold is needed. It is easier to connect through gaps, than to resolve crossings.
Another thing that improves the result a little bit more is to apply a "morphological closing" before using FiberFind. For this, use ProcessGeo > Morphological Operations > Dilate.
First reassign the non-fiber material to pore material (ProcessGeo - Reassign Materials & Material IDs).
Then, dilate the fiber material with the same ID.
Afterwards dilate the pore material with pore material.
Then, the fibers are little bit smoothed without losing to much of their shape.
This helps FiberFind to identify the fibers afterwards.
Best regards,
Janine