Conditional Wasserstein GANs for Image Dehazing

Project Supervisor: Professor Sudipta Mukhopadhyay (IIT Kharagpur)

Achieved state-of-the-art results for single image dehazing by training a conditional Wasserstein GAN using the pix2pix model architecture. The approach combined multiple loss functions (perceptual loss, MSE, L1, and texture loss) and was evaluated on standard fog removal datasets (D-Hazy and O-Haze).

Implementation Details:

• Framework: PyTorch
• Model: Conditional Wasserstein GAN (pix2pix)
• Training datasets: D-Hazy, O-Haze
• Key innovation: Multi-component loss function design for fog removal

Result: Significantly improved dehazing quality compared to traditional methods, with realistic reconstruction of haze-occluded regions.