MEF-GAN

This is the code for "multi-exposure image fusion via generative adversarial networks".

Architecture:

Fused results:

To train:

CUDA_VISIBLE_DEVICES=0,1 python main.py
(2 GPUs are needed. One is for the self-attention block and the other one is for other blocks and the discriminator.)

To test:

CUDA_VISIBLE_DEVICES=0,1 python test_main.py

Tips:

The training dataset is too large to be uploaded and downloaded. It may be more convenient to create your own dataset.
The multi-exposure image pairs can be downloaded here.
The code to create your own training dataset can be found here.
(size_input=144. The channel dimension: 1:3 over-exposed patches, 4:6 under-exposed patches, 7:9 ground-truth patches.)