Light-Up
Image Enhancement
Note: Please search in google for under-exposed or low contrast images before trying the web-app.
Quick Start: Enhance Low light Images -https://brightenhance.herokuapp.com/ Low-end version- https://enhanceimage.herokuapp.com/ [In case of hicupps, please referesh:)]
Losses
https://wandb.ai/vijish/uncategorized/reports/Losses---VmlldzoyNjYwNjc
Generator output (media)
https://wandb.ai/vijish/uncategorized/reports/Output--VmlldzoyNjYwNzA
Table of Contents
About Light-Up
The aim of the project is to enhance under-exposed Images. Before going into technical details I would like to show some pictures.
Example Images
Extremely Dark
Almost NoGAN
The steps are as follows:
- Train the generator with feature loss.
- Train the critic on distinguishing between those outputs and real images.
- Finally, train the generator and critic together in a GAN.
All the useful GAN training here only takes place within a very small window of time(thanks to DeOldify), This helped me do the whole project in Colab. The GAN training took about 25-30 minutes.
Technical Details
DeOldify
Self-Attention Generative Adversarial Network
-Generator is pretrained U-Net
-This has been modified to have spectral normalization along with self attention.
Note: Perceptual Loss (or Feature Loss) based on VGG16--(Thanks to #Fast.ai)
Perceptual Losses for Real-Time Style Transfer and Super-Resolution
Progressive Growing of GANs
Size of the input is progressively Changed and the learning rates are adjusted to make sure that the transitions between sizes happened successfully.
Docker
Clone the repo and navigate to the repo:
git clone https://github.com/vijishmadhavan/Light-Up.git app
cd app/enhance
Build and run the docker image locally:
make run
Navigate to http://localhost:8501 for the app. (Streamlit runs on port 8501 by default)
Shutdown the server:
make stop
Installation Details
This project is built around the wonderful Fast.AI library.
- fastai==1.0.61 (and its dependencies). Please dont install the higher versions
- PyTorch 1.6.0 Please don't install the higher versions
Credits
Project - https://github.com/jantic/DeOldify
Copyright (c) 2018 Jason Antic
License (MIT)-https://github.com/jantic/DeOldify/blob/master/LICENSE