SFA-GAN: Structure-frequency-aware generative adversarial network for underwater image enhancement

Abstract

Influenced by light scattering, absorption and water impurities, the quality of underwater image is so poor that it poses a great challenge to underwater target detection, marine biological research and marine exploration. Thus, significant attention on underwater image enhancement (UIE) has been attracted for producing high quality visuality as if the underwater image was taken in-air without any structure, texture and color loss. To solve this issue, previous work mainly focus on supervised-learning with large amount of paired data, which is more demanding in practical application. Recent Cycle-GAN based UIE break through the dependence on paired data but easily trap in mapping ambiguity. Essentially, two-sided cycle-consistency is a bijection and only focuses on the pixel level, which is too restrictive and can not accurately express underwater scene structure. Besides, high frequencies in reference images tend to be eschewed by generator, making it difficult to synthesize authentic textures and colors of underwater images. We therefore propose a novel unconstrained UIE framework, structure-frequency-aware generative adversarial network (SFA-GAN), which not only accurately preserves the structure of low quality underwater images, but also captures the high frequencies of the reference images under unconstrained settings. Extensive experiments on datasets EUVP, UFO-120 and UIEB demonstrate that the proposed SFA-GAN can achieves state-of-art results on some metrics and produce more clear underwater images without sacrificing model complexity.