Unsupervised Image Translation for Underwater Image Enhancement

Underwater images are often degraded due to the absorption and scattering of light, leading to color distortion, reduced contrast, and blurriness. Enhancing these images is vital for improving visibility in marine applications such as underwater robotics, oceanographic surveys, and archaeological studies. In this paper, we propose a novel unsupervised image translation approach for underwater image enhancement using a Cycle-Consistent Generative Adversarial Network (CycleGAN). Unlike supervised models, our method requires no paired data, addressing a major limitation in underwater datasets. We enhance the baseline CycleGAN framework by introducing perceptual loss and structural similarity index (SSIM) loss to preserve semantic and structural details. Experimental results on multiple underwater datasets demonstrate significant improvement in image quality, outperforming several state-of-the-art methods both quantitatively and qualitatively.