Imaginary-to-Real Generative Network for Marine Organism Image Extrapolation

Marine organism research plays an extremely important role in ocean science. Underwater computer vision is a main measure to observe organism. In this work, we devote on drawing the entire organism image by a visible part. Inspired by the human ability to “imagine-draw” an image, we propose a novel Imaginary-to-Real Generative Network, ImReNet for short, facing the task's challenges and overcoming the semantic inaccuracies and background distortions problems. We devise a method of image generation with an imaginary output as guidance, involving a novel underwater-color loss to deal with the similar chromatic characteristics of underwater images. Experiments conducted on datasets covering both Fish and Plankton present that our method achieves good performance on image extrapolation. Our code will be made publicly available.