Deep Facial Reconstruction

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2022-01-24

Department

Major/Subject

Autonomous Systems

Mcode

ELEC3055

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

56

Series

Abstract

Everyone wants their images to look as good as possible when they post on social media. It is not always possible to retake a picture or to manually edit small mistakes afterwards. This paper investigates the viability of using deep learning to perform this task for a type of edit. More specifically, we want to “photoshop” a face onto a subject wearing a facemask. To explore this possibility, we examined techniques and methodologies from facial attribute removal and image completion. Then, we structurally test and implement the most promising ideas. The result is a CycleGAN model capable of reconstructing a realistic looking face. The model matches performance with state-of-the-art image completion and outperforms models in facial attribute removal.

Description

Supervisor

Kannala, Juho

Thesis advisor

Hadid, Abdenour

Keywords

computer vision, image completion, generative adversarial network, GAN, CycleGAN, DMFN

Other note

Citation