University of Twente Student Theses


Training Facial Recognition with Synthetic Faces

Vine, P.H. (2021) Training Facial Recognition with Synthetic Faces.

[img] PDF
Abstract:The effective generation of synthetic faces may be useful for improving facial recognition training datasets. This work explores methods for generating synthetic faces and trained a generative network to synthesize front facing facial images of existing identities with different attributes as well as of completely new identities. The identities of the synthetic faces were evaluated using 3 pretrained facial recognition systems. Facial recognition networks were trained to compare the performance of training with the synthetic faces and real faces. The ability to use the synthetic faces for data augmentation was also evaluated. It was found that the mean equal error rate (EER) increased from 2.21% when using the real facial images to 5.27% when training with completely synthetic faces of new identities. When using the synthetic faces for data augmentation, the new identities could improve the mean EER. However, this improvement is not guaranteed with some training datasets leading to higher mean EER after training with more synthetic faces. There is clearly still a difference between the synthetic faces generated and real faces. Understanding what is still missing in the synthesized faces would be valuable research to more effectively enable training facial recognition with only synthetic faces.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Electrical Engineering MSc (60353)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page