Using GANs to reconstruct and enhance the quality of underwater hazed images

This model takes inspiration from the work proposed with FUnIE-GAN, with a few modifications:

• A perceptual colour distance function is used to enhance the quality of reconstructed colours.

• The generator architecture was changed to one inspired by the one used in EnlightenGAN, so instead of using a nearest-neighbour upsampling layer the generator is fully-convolutional.

• The generator loss was changed from MSE to MAE to improve its stability. Additionally, one-sided label smoothing was used to improve the generator's adversarial defense.

The generator architecture is composed by 8 upsampling and 8 downsampling blocks.

Additionally, a smaller version of the generator is also available, which is identical to this one but with only 5 and 5 blocks instead.

Some dehazing examples:

References:

• FUnIE-GAN github, paper

• EnlightenGAN paper

• Colour perceptual distance Definition, implementation