I am looking for light weight (embedded) model for face super resolution.

Despite detecting all faces in images taken from an indoor environment, I have a problem verifying tiny faces (the embedding vectors for tiny faces do not provide sufficient information for face verification).

I want to use super resolution to enhance the resolution of the cropped face after face detection, then I want to get the embedding vector for this enhanced cropped face.

Which models of super resolution are efficient for this process?, considering that super resolution model needed to be lightweight model and with efficient execution time

(GANs model are not very good for face detection and verification because of their execution time, model size and some artificial results (like GFPGAN and PSFRGAN)

For face detection and verification, I use InsightFace.

Please guide me with your ideas and experiences. Thanks very much.

New contributor
David is a new contributor to this site. Take care in asking for clarification, commenting, and answering. Check out our Code of Conduct.
  • 1
    $\begingroup$ Just checking: You are aware that all super-resolution models invent feasible filler data based on their training sets? There are no non-generative super-resolution models (although plenty are non-random, they simply generate the mode of the prediction - it is still generative). It is a very risky step to add such a generative model into an identifer pipeline. You will get a higher false match rate. $\endgroup$ 2 days ago


You must log in to answer this question.

Browse other questions tagged .