University of Hertfordshire

By the same authors

  • Daniel Sáez-Trigueros (Inventor)
  • Li Meng (Inventor)
  • Margaret Hartnett (Inventor)
View graph of relations
Original languageEnglish
Patent numberEP 3 428 843 A1
IPCG06K 9/00 (2006.01)
Priority date14/07/17
Publication statusPublished - 16 Jan 2019

Abstract

A method of identifying at least one face region having a predominant contribution to face recognition, comprising (i) obtaining a set of face images, (ii) using a face-recognition model to predict a non-occluded identity class for a first face image, (iii) placing an occluder at a first location of the first face image, (iv) using the face-recognition model to predict an occluded identity class for the occluded first face image, (v) comparing the non-occluded identity class with the occluded identity class to generate an occlusion result, (vi) repeating steps (iii) to (v) for a next location of a plurality of locations of the first face image to generate an occlusion map comprising the occlusion results at each of the locations of the first face image, (vii) repeating steps (ii) to (vi) for a next face image of the set of face images to generate an occlusion map for each of the set of face images, (viii) combining the occlusion maps for each of the set of face images to generate an aggregate occlusion map, and (ix) using the aggregate occlusion map to identify the at least one face region having a predominant contribution to face recognition. The invention further comprises a computer program for execution of the method of identifying at least one face region having a predominant contribution to face recognition and a system for identifying at least one face region having a predominant contribution to face recognition. The invention further comprises a method of augmenting a set of training face images for a face-recognition model comprising using the at least one face region having a predominant contribution to face recognition and a method of training a face-recognition model to recognise faces using the augmented set of training face images.

ID: 15637672