Recovering Super-Resolution Generative Adversarial Network for Underwater Images

In this paper, we propose an end-to-end Recovering Super-Resolution Generative Adversarial Network (RSRGAN) to automatically learn super-resolution underwater images. RSRGAN mainly includes two parts. The first part is a Recovering GAN, aiming at color correction and removing noise in the images. The generator of Recovering GAN is based on an encoder-decoder network with self-attention on the global feature. The second part is a Super-Resolution GAN, which adopts the residual-in-residual dense block in its generator, to add details onto the results fed from the Recovering GAN. Both qualitative and quantitative experimental results show the advantage of RSRGAN over the state-of-the-art approaches for underwater image super-resolution.