Self-Adversarial Generative Adversarial Network for Underwater Image Enhancement

Generative adversarial network (GAN)-based underwater image enhancement methods improve image quality by encoding a raw image and an adversarial image with the corresponding high-quality version. Although these methods have recently attracted significant attention, the lack of referred clear underwater imagery forces GAN-based underwater image enhancement models to be trained with synthetic or enhanced underwater images, limiting their applicability and performance. This article proposes a novel self-adversarial GAN (SA-GAN) to enhance underwater images by referring to paired raw and high-quality natural images. Specifically, a self-adversarial mode is designed that attaches a further constraint to the generation procedure. By applying two pairwise image quality discriminators, the generators are supervised with a stronger decision boundary to generate better quality than the high-quality natural image and the last-generated image. This is a novel settlement of the limitation caused by the adversary system with the synthetic or enhanced underwater images, realizing the quality transfer from natural images to distorted underwater images. Several experiments on real underwater images and two commonly used underwater image data sets demonstrate that the proposed method subjectively and objectively performs better than current methods in restoring the coloration of underwater images.