README.md

Enhancing Images with Uneven Illumination using Ensemble Learning (EIUIE)

This repository contains the codebase for the EIUIE approach to enhancing images with uneven illumination.

Usage

To enhance a single image, the following cammand can be used:

python eiuie/main.py single --file=<PATH_TO_FILE> --method=unsharp_masking
python eiuie/main.py single --file=<PATH_TO_FILE> --method=retinex
python eiuie/main.py single --file=<PATH_TO_FILE> --method=homomorphic_filtering
python eiuie/main.py single --file=<PATH_TO_FILE> --method=fusion

In order to use the fusion model, its parameters have to be trained first.

Training

First, the LOL-dataset has to be downloaded, extracted, and moved to data/lol_dataset/.

From this dataset training samples will be generated. This happens in two steps: In the first step, the three base methods (unsharp_masking, retinex, homomorphic_filtering) are applied to all images in the LOL-dataset. Results will be stored in data/intermediate_images/

python eiuie/main.py batch_process

Afterwards, the training dataset consisting of all pixel samples can be generated to data/pixel_dataset/ using:

python eiuie/main.py prepare_dataset

Training can then be conducted, with the best model parameters (best_model.pt) as well as further checkpoints being saved to data/checkpoints/, using:

python eiuie/main.py train

The learnt parameters can be pretty printed using:

python eiuie/main.py ppparams

Finally, training can be evaluated on a custon evaluation dataset (data/eval_dataset/input/) with the following command. Output images and scores for different evaluation metrics can be found afterwards in data/eval_dataset.

python eiuie/main.py eval