Abstract:

Researchers have been focusing on uncovering underwater treasures by overcoming the obstacles of poor quality underwater images.. The onset of deep learning methods and the various acquisitions of underwater images paved the way for a lot of explorations. In this paper, architecture based on GAN is proposed to enhance the characteristics of the underwater image by preserving the structure and content of the image. Experiments are executed by utilizing the publicly available UFO-120 and UIEB datasets which include both real undersea images and their corresponding reference images. To boost the performance of the architecture, ℒ₁ and content-based loss are combined with ℒ_cGAN. The final enhanced image provides an appealing result in qualitative evaluation whereas the results obtained from PSNR, SSIM, and UIQM metrics demonstrate that the suggested strategy produces improved results when compared with the most recent techniques.

Key words: generative adversarial network, generator, discriminator, loss function