ISSN 2079-3537      

 
 
 
                                                                                                                                                                                                                                                                                                                                                                                                                                                                             





Scientific Visualization, 2022, volume 14, number 2, pages 1 - 17, DOI: 10.26583/sv.14.2.01

Method for Generating Synthetic Images of
Masked Human Faces

Authors: M.A. Letenkov1,A,B, R.N. Iakovlev2,A,B, M.V. Markitantov3,A,B, D.A. Ryumin4,A,B, A.I. Saveliev5,A,B, A.A. Karpov6,A,B

A St. Petersburg Federal Research Center of the Russian Academy of Sciences (SPC RAS), Russia

B St. Petersburg Institute for Informatics and Automation of the Russian Academy of Sciences, Russia

1 ORCID: 0000-0001-5745-5354, o1prime@yandex.ru

2 ORCID: 0000-0002-6721-9707, iakovlev.r@mail.ru

3 ORCID: 0000-0001-7987-1025, m.markitantov@yandex.ru

4 ORCID: 0000-0002-7935-0569, ryumin.d@iias.spb.su

5 ORCID: 0000-0003-1851-2699, saveliev.ais@yandex.ru

6 ORCID: 0000-0003-3424-652X, karpov@iias.spb.su

 

Abstract

This study is devoted to the topical problem of generating synthetic images of human faces. The paper presents a new method for generating images of human faces in protective masks. The proposed method is based on the combined use of a neural network method for detecting three-dimensional facial landmarks (3D-FAN) and three-dimensional modeling tools. Approbation and quality assessment of the proposed method was conducted on a test dataset, which includes 3836 images. The dataset included human faces images of different gender and age, taken at different distances and at various angles relatively to the camera lens. To assess generation results, the method of multi-criteria assessment was used with the involvement of an expert group. For each generated image final scores were formed by averaging the obtained ratings, both by criteria and by experts. During the experiment, the developed method has demonstrated a high and stable quality for the following ranges of face orientations [-20; +55], [-60; +60] and [-70; +80] along the OX, OY and OZ axes, respectively. The final proportion of correctly generated images of masked human faces turned out to be 95.9%.

 

Keywords: synthetic face generation, synthetic visual data corpus, masked face generation, 3D modeling, 3D-FAN, Blender.