Attempting to fine tune WEKA Segmentation Settings

62A SEM-03 7500x Fine-1.tif (284.1 KB) 04C SEM-03 7500x Coarse-1.tif (283.9 KB) 04D SEM-06 7500x Fine-1.tif (281.7 KB) #1 SEM-02 7500X Dend-1.tif (279.8 KB) TSEM-02 7500X Dend-1.tif (278.7 KB)

These are a typical cross section of the images what I will be analyzing, needing to measure the darker blocks or blobs.

I have got WEKA segmentation working, but am unsure as to which of these settings will improve preformance and which will just waste time.


It’s always a balance between selecting features and time for training… but this is a balance only you can really determine. Try to use the minimal number of features to get the segmentation that is most appropriate for your biology. You can read up on the different features here and try to narrow down those that are most applicable in your case.

But you also said:

So are you happy with the segmentation results you are getting? Is it really taking too long to generate?

Also keep in mind that the more pixels you select for each class - the longer it will take as well… so perhaps you can add that to the mix to balance as well?

1 Like

I am currently having issues with the first two or three trainings having nearby gamma prime blended together and having the lighter sections of large gamma prime seen as base material, when training on the default settings shown. While I have read the whole Trainable WEKA Segmentation page, I am unsure as to which of those features and perhaps what alterations to the random forest would even be helpful for my use given that I am not well versed in image analysis, image manipulation or the like.

Also of minor note, this is metallurgic engineering not biology, but that really only matters on the end user side of the analysis.

Is there anyway of getting a comparison between the frequency a setting’s data is used with the accuracy it delivers or it’s run time?


For both posts - I would defer to the developer of this tool… @iarganda

Regarding run time vs number of features, you have to check the log window and see how many image features are produced based on your settings. By default, there are probably too many. Some of them are texture descriptors, some of them are border detectors, etc. You really should try to think what is more important to classify a pixel based on its neighbors.

One other thing to take into account is the use of normalization as a pre-processing step in your images, so all of them present more or less the same histogram and training a model on some of them could then be applied to others. Have a look at this post.