Generating Synthetic Disguised Faces with Cycle-Consistency Loss and an Automated Filtering Algorithm

• Published in: Mathematics (Basel) Vol. 10; no. 1; p. 4
• Main Authors: Ahmad, Mobeen, Cheema, Usman, Abdullah, Muhammad, Moon, Seungbin, Han, Dongil
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.01.2022
• Subjects: Algorithms; Automation; Cameras; Consistency; CycleGAN; disguised face; Face recognition; Facial recognition technology; Generative adversarial networks; Image filters; Mathematics; Methods; Object recognition; Spoofing; style transfer; Synthesis; synthetic database; synthetic faces; Training
• ISSN: 2227-7390; 2227-7390
• DOI: 10.3390/math10010004

Applications for facial recognition have eased the process of personal identification. However, there are increasing concerns about the performance of these systems against the challenges of presentation attacks, spoofing, and disguises. One of the reasons for the lack of a robustness of facial recognition algorithms in these challenges is the limited amount of suitable training data. This lack of training data can be addressed by creating a database with the subjects having several disguises, but this is an expensive process. Another approach is to use generative adversarial networks to synthesize facial images with the required disguise add-ons. In this paper, we present a synthetic disguised face database for the training and evaluation of robust facial recognition algorithms. Furthermore, we present a methodology for generating synthetic facial images for the desired disguise add-ons. Cycle-consistency loss is used to generate facial images with disguises, e.g., fake beards, makeup, and glasses, from normal face images. Additionally, an automated filtering scheme is presented for automated data filtering from the synthesized faces. Finally, facial recognition experiments are performed on the proposed synthetic data to show the efficacy of the proposed methodology and the presented database. Training on the proposed database achieves an improvement in the rank-1 recognition rate (68.3%), over a model trained on the original nondisguised face images.