A Light SRGAN for Low Rsolution and High Latency Images

The model, derived from original SRGAN is supposed to work on much smaller images and with a much higher noise value. The target of the paper was to create a Model that is fast, light and trained on a much smaller dataset and time.

The paper was reviewed at 2 places before being published at ICACDS'21 along with being published in Springer CCIS Journal.


Abstract

In the past few years Single Image Super-Resolution(SISR) has been one of the most researched topics in the field of AI. SuperResolution Generative Adversarial Nets in short SRGAN paved the way to achieve Super-Resolution(SR) of images while hallucinating a lot of details. Deriving from the main components from SRGAN, i.e. Architecture, Loss and Adversarial nature, we have refined a model that works for very small images, and tries to make out as much information as possible in a short amount of time. The main things being focused are to create a fast Generator which also tries to keep a good SSIM score with the ground truth images, tries to recover as much of the information from relative pixels and also gets close enough to benchmark performance with as limited resources as possible. The core objective of having a simple, fast and light model, is not only to enlarge images but fill in as many missing details as it can from simple pixels, to fully defined and distinct features within that image that might have double or quadruple resolution than the Low-Resolution Images.

Network Architecture

Generator

The Full Size Generator can Be found Here

Discriminator


The Full Size Discriminator can Be found Here


Dataset Used

We have used the following dataset for training:

COIL-100

Div2k-Training Split

And the following dataset as test:

Set-5

Set-14

The code are not provided in the exact manner because of copyright, but they can be easily recreated in a TF.2x Env.

Feel free to contact if you have any queries.