Super resolution faces captured from video clips

Algorith detecting, recognizing and enhancing resolutions of captured faces faces from viedo clips.

1. Super Resolution Generative Adversarial Network

First part is SRGAN model - a seminal work that is capable of generating realistic textures during single image super-resolution - which is being trained on "CelebA Dataset"

Presented SRGAN is an implementation of "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network"

1.1 Architecture

SRGAN is a generative adversarial network for single image super-resolution. It uses a perceptual loss function which consists of an adversarial loss and a content loss. The adversarial loss pushes the solution to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved images and original photo-realistic images.

Architecture of Generator and Discriminator Network with corresponding kernel size (k), number of feature maps (n) and stride (s) indicated for each convolutional layer.

1.2 Inputs and oututs

As inputs goes low-resolution images and high-resolution images and the generator based on them as outputs generate super-resoultion images from low-resolution ones. Below are presented some samples form training process:

Results after 500 epchos

Results after 20 000 epchos

Results after 50 000 epchos

2. Face detection and recognition

Second part is detecting, recognizing known faces and improving resolution by using generator from SRGAN on viedo clips.

Face detection and recognition.

Super-resolution face enhancment