MSSCE-GAN: Multi-Scale Structural and Color Enhanced Generative Adversarial Network for Unpaired Underwater Image Enhancement

Enhancing underwater images is pivotal for applications like underwater exploration. While most existing methods require paired underwater and reference images for training, acquiring such reference images is difficult. Current techniques often suffer from information loss, resulting in blurred details and lack generalizability. This paper introduces an unpaired underwater image enhancement method using the Multi-Scale Structural and Color Enhanced Generative Adversarial Network (MSSCE-GAN). We also introduce a detail feature recovery module and an attention enhancement module, targeting multiple distortions in underwater imagery. Importantly, our method can produce superior enhanced images without paired training data. Experimental evaluations demonstrate our approach outperforms existing techniques in both effectiveness and generalizability on multiple underwater image datasets.