Patent · US Active

Geometrically constrained, unsupervised training of convolutional autoencoders for extraction of eye landmarks

US12093835B2 · kind B2 · utility

0Cited by
0References
16Claims
0Family size

Inventors

Key dates

Filing dateOct 23, 2022
Grant dateSep 17, 2024
Priority date
Expiry dateOct 23, 2042

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG06T2207/20132
  • WIPO fieldComputer technology
  • WIPO sectorElectrical engineering

Abstract

Unsupervised, deep learning of eye-landmarks in a user-specific eyes' image data by capturing an unlabeled image comprising an eye region of a user, using an initial geometrically regularized loss function, training a plurality of convolutional autoencoders on the unlabeled image comprising the eye region of the user to recover a plurality of user-specific eye landmarks, training a convolutional neural network for autoencoded landmarks-based recovery from the unlabeled image, and where the initial geometrically regularized loss function is represented by the formula LAE=λreconLrecon+λconcLconc+λsepLsep+λeqvLeqv where LAE is total AutoEncoder Loss, λreconLrecon is λ-weighted reconstruction loss, λconcLconce is λ-weighted concentration loss, λsepLsep is λ-weighted separation loss, and λeqvLeqv is λ-weighted equivalence loss.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.