Underwater image enhancement using CycleGAN

Abstract:
Underwater visual environment perception is an important issue for autonomous motion and operation of underwater robots. Due to the absorption of light by water and the scattering of light by particles in the underwater environment, the underwater image has low contrast and blurry edges. In this paper, we propose a method to capture image domain features and convert them into another image domain without any paired training samples. Cyclic consistent loss is used to eliminate pairing data during training. The model can convert from one domain to another without one-to-one mapping between the source domain and the target domain. For providing a feasible solution for deploying real-time GAN on resource constrained devices, a method based on GAN compression is proposed to generate high undistorted images with low computational requirements. This method can be applied to many other tasks, such as image enhancement, image coloring, style transfer.