Underwater-GAN: Underwater Image Restoration via Conditional Generative Adversarial Network

Underwater image restoration is still a challenging task until now, because underwater images are degenerated due to the complex underwater imaging environment and poor light condition. The image degeneration includes color distortion, low contrast, and blur. In this paper, we propose Underwater-GAN, a conditional generative adversarial network for underwater image restoration. In Underwater-GAN, we use Wasserstein GAN with gradient penalty term as the backbone network. We design the loss function as the sum of the loss of generative adversarial network and the perceptual loss. In the discriminator of Underwater-GAN, we use a convolution patchGAN classifier to learn a structural loss instead of the image-level loss or pixel-wise loss. Moreover, we construct an underwater image dataset by simulating to generate underwater images according to the underwater imaging model. We train our model with these simulated underwater dataset. The results of our experiments show that the proposed method produces better visual qualitative and quantitative indicators than existing methods.