Conditional Wasserstein Generative Adversarial Network With W-Net and Wasserstein-L1-SSIM Loss Function for Single Image Dehazing
Bijaylaxmi Das, Hemanta Kumar Sahu, Ashish Kumar Padhan, Sudipta Mukhopadhyay · IEEE Access · 2026
Generative Adversarial Networks (GANs) have been widely used to solve different image translation problems. This article proposes an exploration of the Conditional Wasserstein Generative Adversarial Network (cWGAN) building blocks to improve its performance. The improvements in the generator, discriminator, and/or loss functions can improve the training of the network and the quality of the outputs. A W-Net has been introduced as the network’s generator to improve performance. A new combination of Wasserstein, L1 loss, and SSIM (WaLSS) loss functions is proposed for network training to improve the quality of generated images. Haze and other atmospheric conditions reduce vision, posing challenges for computer vision applications. The lack of prior knowledge presents difficulties in single-image de-hazing. The proposed refined Conditional Wasserstein Generative Adversarial Network with W-Net and Wasserstein-L1-SSIM Loss Function (cWGAN-WaLSS) is a complete network that requires no prior knowledge and no further enhancements. Hence, cWGAN-WaLSS is a good candidate for single-image de-hazing. Extensive testing with real and simulated hazy images shows that the proposed cWGAN-WaLSS network demonstrates consistent and competitive performance relative to existing cWGAN-based and other dehazing methods.