Feature Engineering - translational VARIANT and rotational invariant



Hey guys,

there are many texture features for classification purposes available. Many of them are invariant against translation, rotation and scaling (TRS invariant).

However, I like the last two properties but for my recent project it would be better if they would be variant against translation.

Are there any ideas how to transform TRS invariant feature into an RS invariant feature?



Hello @twagner,

Are you doing pixel or image classification? Which features are you interested on?


Hi @iarganda,
I’m doing image classification in very noise (SNR << 1) and very low contrast cryo em images. I try to classify small image patches into “contains particle” and “contains no particle”

I’m currently using a set of features like

  • Entropy, Energy, Contrast etc. of the co-occurence matrix
  • Variations of local binary pattern
  • Entropy, Skewness, Kurtosis, Modulus of the Histogram
  • Hu moments
  • Various descriptors of several textbooks.

Unfortunately, I’m not in office and therefore have not access to my code to give you a complete list.



I see, what about using regular uniform Local Binary Pattern features? Aren’t they rotation invariant by translational variant? I would assume that most features based on blocks would make the trick, no?


Local Binary Pattern basically counting geometric features of an greyscale image. If we compare the LBP of a greyscale particle in center of a box with one in the lower left of the box but with the same geometry then I don’t see a reason that the distribution of LBP would change. Or do I miss something? Furthermore LBP are very noise sensitive :-/


I see, so you want to identify your images based on the location of particles on it, is that right? Can’t you use blocks which are smaller than your particles then? That way the blocks would identify the location as well.

Yes, but they are robust to contrast changes. I guess you could also use HOG features and play with the block size depending on your particle sizes.