Face Image Super Resolution using a Generative Adversarial Network

Traditional image super resolution centered around purely mathematical models are capable of creating gradient based textures, but fail to render the specific lineaments that would be expected in a realistically upscaled image. This is especially problematic in scenarios involving images of subjects...

Full description

Saved in:
Bibliographic Details
Published in2021 Smart Technologies, Communication and Robotics (STCR) pp. 1 - 8
Main Authors Varma, Kavya, Reddy, G. Satyabhama, Subramanyam, Natarajan
Format Conference Proceeding
LanguageEnglish
Published IEEE 09.10.2021
Subjects
Deep learning
Face recognition
Facial Enhancement
Feature extraction
Generative adversarial networks
Image recognition
Image resolution
Image Super Resolution
Mathematical models
SRGAN
Online AccessGet full text

Cover

More Information
Summary:Traditional image super resolution centered around purely mathematical models are capable of creating gradient based textures, but fail to render the specific lineaments that would be expected in a realistically upscaled image. This is especially problematic in scenarios involving images of subjects whose recognition is reliant on the presence of specific characteristics, for example, faces. In this paper, we describe a deep learning model that is capable of generating an 8x upscaled photo from a low resolution image of a face. The underlying model is based on the SRGAN architecture that deviates from the conventional GAN approach of the adversarial back and forth between generator and discriminator by incorporating an added content loss whose value is dependent on the detection of the natural features in the generated image by a pre-trained VGG model. The model is trained on the Celeb Faces Attributes dataset with over 1,00,000 data points and can produce upscaled images that are realistic with a coherent presence of the natural attributes of a face.
DOI:10.1109/STCR51658.2021.9588816