I am working on a project where I classify tiny moving particles into a few classes (fibers, hairs, glass shards, bubbles). The particles are only a few pixels large and are observed in a few frames as the particle moves and rotates. As the particles are so tiny, it is expected that classification with a CNN or observed particle characteristics on singular observations do not yield a satisfactory accuracy. A voting system can be used to classify the particle with multiple observations but it might be possible to achieve a better result.

I am looking for a method to combine the different observations of a particle to extract new relevant features. For example, if the shape of the particle varies over the different frames, this could indicate that the particle is not symmetrical.

Do you know any methods that perform such multi-observation classification methods? Thanks a lot!

  • $\begingroup$ I have found one method that does such a thing: paper. It fuses the values of a CNN at different depths in the network. I will look at this method. $\endgroup$ Apr 20 at 13:25


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.